fazlullah khan mian ahmad jan muhammad alam editors

Applications of Intelligent Technologies in Healthcare

Fazlullah KhanMian Ahmad JanMuhammad Alam Editors

EAI/Springer Innovations in Communication and Computing

EAI/Springer Innovations in Communicationand Computing

Series editorImrich Chlamtac, CreateNet, Trento, Italy

Editor’s Note

The impact of information technologies is creating a new world yet not fullyunderstood. The extent and speed of economic, life style and social changesalready perceived in everyday life is hard to estimate without understanding thetechnological driving forces behind it. This series presents contributed volumesfeaturing the latest research and development in the various information engineeringtechnologies that play a key role in this process.

The range of topics, focusing primarily on communications and computing engi-neering include, but are not limited to, wireless networks; mobile communication;design and learning; gaming; interaction; e-health and pervasive healthcare; energymanagement; smart grids; internet of things; cognitive radio networks; computation;cloud computing; ubiquitous connectivity, and in mode general smart living, smartcities, Internet of Things and more. The series publishes a combination of expandedpapers selected from hosted and sponsored European Alliance for Innovation (EAI)conferences that present cutting edge, global research as well as provide newperspectives on traditional related engineering fields. This content, complementedwith open calls for contribution of book titles and individual chapters, togethermaintain Springer’s and EAI’s high standards of academic excellence. The audi-ence for the books consists of researchers, industry professionals, advanced levelstudents as well as practitioners in related fields of activity include information andcommunication specialists, security experts, economists, urban planners, doctors,and in general representatives in all those walks of life affected ad contributing tothe information revolution.

About EAI

EAI is a grassroots member organization initiated through cooperation betweenbusinesses, public, private and government organizations to address the globalchallenges of Europe’s future competitiveness and link the European Researchcommunity with its counterparts around the globe. EAI reaches out to hundreds ofthousands of individual subscribers on all continents and collaborates with an insti-tutional member base including Fortune 500 companies, government organizations,and educational institutions, provide a free research and innovation platform.

Through its open free membership model EAI promotes a new research and inno-vation culture based on collaboration, connectivity and recognition of excellence bycommunity.

More information about this series at http://www.springer.com/series/15427

http://www.springer.com/series/15427

Fazlullah Khan • Mian Ahmad JanMuhammad AlamEditors

Applications of IntelligentTechnologies in Healthcare

123

EditorsFazlullah KhanAbdul Wali Khan University MardanMarden, Pakistan

Muhammad AlamXi’an Jiaotong-Liverpool UniversitySuzhou, China

Mian Ahmad JanAbdul Wali Khan University MardanMarden, Pakistan

ISSN 2522-8595 ISSN 2522-8609 (electronic)EAI/Springer Innovations in Communication and ComputingISBN 978-3-319-96138-5 ISBN 978-3-319-96139-2 (eBook)https://doi.org/10.1007/978-3-319-96139-2

Library of Congress Control Number: 2018959865

© Springer Nature Switzerland AG. 2019This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part ofthe material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,broadcasting, reproduction on microfilms or in any other physical way, and transmission or informationstorage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodologynow known or hereafter developed.The use of general descriptive names, registered names, trademarks, service marks, etc. in this publicationdoes not imply, even in the absence of a specific statement, that such names are exempt from the relevantprotective laws and regulations and therefore free for general use.The publisher, the authors, and the editors are safe to assume that the advice and information in this bookare believed to be true and accurate at the date of publication. Neither the publisher nor the authors orthe editors give a warranty, express or implied, with respect to the material contained herein or for anyerrors or omissions that may have been made. The publisher remains neutral with regard to jurisdictionalclaims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG.The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

https://doi.org/10.1007/978-3-319-96139-2

To our wonderful families, especially thekids.

Fazlullah Khan, Mian Ahmad Jan,Muhammad Alam

Foreword

The world has seen significant growth in the advancement of wireless technologiesfrom communication to health control and monitoring applications. In this regard,in last two decades, numerous innovative wireless technologies such as mobilephones, Wi-Fi, high-speed Internet, smart houses, smart cities, smart hospitals,underwater networks and body area networks, etc. have emerged, and many arestill under consideration. The academia and industrial research groups are facinghuge demands from the users to develop wireless applications to fulfill theirneeds of high data rates and error-free communication, maximizing battery lifetime and bandwidth efficiency. Therefore, this book mainly focuses on presentingthe new emerging paradigms in wireless communication for mitigating and/orinvestigating health problems, specifically, characterizing the concept of Internetof Things, Big Data, and cloud computing in detecting various deadly diseases,such as heart diseases, asthma, tuberculosis. For example, the Internet of Thingshave been studied in building smart hospital rooms and beds for the sake ofelectronically/remotely monitoring patients through wireless devices. This will helpthe doctors to visit only critical patients. Moreover, with the help of body areanetwork, the patient’s details such as blood pressure, heartbeat, and temperature areremotely monitored by the doctor, and therefore the patient will visit the hospitalonly in the case of emergency.

To this point, it looks so fantastic to connect everything; however, one can avoidthe data storage problem of connecting hundreds of wireless applications. Therefore,this book also focuses on how to optimally store data considering traditional andcloud storage. The cloud computing plays an important role in storing huge amountof data that has been processed by various cloud servers which lower not only theburden of storage but also the processing power of the traditional systems. To doso, many efficient machine learning algorithms are used for processing this hugeamount of data.

Department of Computer Science Ateeq Ur RehmanAbdul Wali Khan University MardanMardan, Pakistan

vii

Preface

This book provides detailed insight information about the use of health informaticsand emerging technologies for the well-being of patients. The latest technologicalgrowth in the healthcare sector has revolutionized the facilities provided at eachhospital and pharmacy. Previously, the health industry relied on a paper-basedsystem to organize, store, interpret, and integrate patient’s records and medicalinformation. However, with the informatics industry booming and allowing fornew electronic technology and information systems, practitioners now find this datastored in convenient coded computer systems. For example, all the patient’s recordsare maintained electronically at a reduced cost with fewer chances of errors. Patientsare no longer required to remember their medical past amid the stress of emergencysituations. Upon entering a hospital or a facility, the medical staffs already havesufficient information available in the electronic health record system that can beused for reference and checking previous history of the patient. As a result, thepatients need not to fill the admission form upon entering the facility each time.Apart from reduced paper work, the patients are no longer required to conductvarious tests each time for diagnosis of various illnesses. The medical staffs alreadyhave a previous history of the patients available in front of their monitors.

This book covers a diverse range of topics that are all related with healthcare.Each chapter in this book emphasizes on the use of informatics in healthcare.In general, each chapter uses various emerging technologies such as Internet ofThings (IoT), Big Data, cloud computing, Wireless Body Area Networks (WBANs),for various health-related illness, such as tuberculosis, heart diseases, asthma, andvarious epidemic outbreaks. The advancements in the IoT have enabled hospitals tostart implementing “smart beds” that are capable to detect when they are occupiedby patients and when patients are attempting to get up. The use of WBANs enablesthe practitioners to remotely monitor the patients. As result, the patients are nolonger required to visit the hospital regularly and also are not required to beadmitted. This allows the hospital to allocate beds to more critical patients who arein urgent need of hospitalization and emergency services. Various symptoms of thepatients are monitored constantly at home and transmitted to hospital and medical

ix

x Preface

staff on regular basis. The technological growth of cloud computing has a pivotalrole in such applications. Effective and efficient services of the cloud are in direneed in such circumstances. Today, the data centers of the cloud have the ability tostore an enormous amount of data that are remotely processed on virtual servers.As a result, the computational burden on the patients’ monitoring devices and thehospital workstations is greatly reduced. In fact, the cloud servers have the abilityto take care of all the computational and resource-intensive operation involved inprocessing, storage, and transmission of any health-related data. The data generatedfrom various medical appliances are enormous, and as such the term “Big Data”plays a pivotal role. Various machine learning algorithms are required to processsuch huge volume of data.

Each chapter of this book is dedicated to the aforementioned tasks. Excepttechnical terms, the readability and flow of concepts make this book an easy togo for any healthcare practitioners. These concepts have the ability to enhance thecare of elderly people and patients even better in healthcare industry.

Mardan, Pakistan Fazlullah KhanMardan, Pakistan Mian Ahmad JanSuzhou, China Muhammad Alam

Acknowledgment

We would like to express our special appreciations and thanks to the Vice Chancellorof Abdul Wali Khan University Mardan and the Patron in Chief of the University ofLahore, Pakistan, for their enduring support.

We would also like to thank authors, reviewers, and committee members.Without their contributions, this book would not have been possible. Our specialgratitude to our friends and colleagues for their technical support, especially Dr.Rodziah Binti Attan, Dr. Abid Yahya, Dr. Nabeel Younus Khan, Dr. Khalid Hussain,Dr. Aurangzeb Khan, Dr. Saeed Islam, Dr. Sajjad Khan, Dr. Mukhtaj Khan, andDr. Ateeq Ur Rehman.

Finally we would like to thank you all who help and support us for everything,and we can’t thank you enough for encouraging us throughout this experience.

Fazlullah Khan, Mian Ahmad Jan, Muhammad Alam

Fazlullah KhanMian Ahmad Jan

Muhammad Alam

xi

Contents

Quality Assessment and Classification of Heart SoundsUsing PCG Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1Qurat-ul-ain Mubarak, Muhammad Usman Akram, Arslan Shaukat,and Aneeqa Ramazan

Classification of Normal Heart Beats Using Spectraland Nonspectral Features for Phonocardiography Signals . . . . . . . . . . . . . . . . . . 13Shahid Ismail Malik and Imran Siddiqi

Segmentation of Chest Radiographs for Tuberculosis ScreeningUsing Kernel Mapping and Graph Cuts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25Ayesha Fatima, Anam Tariq, Mahmood Akhtar, and Hira Zahid

Survey Analysis of Automatic Detection and Grading of CataractUsing Different Imaging Modalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35Isma Shaheen and Anam Tariq

A Privacy Risk Assessment for the Internet of Things in Healthcare . . . . . . 47Mahmoud Elkhodr, Belal Alsinglawi, and Mohammad Alshehri

Parallel Computation on Large-Scale DNA Sequences . . . . . . . . . . . . . . . . . . . . . . 55Abdul Majid, Mukhtaj Khan, Mushtaq Khan, Jamil Ahmad, Maozhen Li,and Rehan Zafar Paracha

Augmented and Virtual Reality in Mobile Fitness Applications:A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67Ryan Alturki and Valerie Gay

Cloud-Assisted IoT-Based Smart Respiratory Monitoring Systemfor Asthma Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77Syed Tauhid Ullah Shah, Faizan Badshah, Faheem Dad, Nouman Amin,and Mian Ahmad Jan

xiii

xiv Contents

Blood Cell Counting and Segmentation Using Image ProcessingTechniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87Ayesha Hoor Chaudhary, Javeria Ikhlaq, Muhammad Aksam Iftikhar,and Maham Alvi

Smart Assist: Smartphone-Based Drug Compliance for ElderlyPeople and People with Special Needs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99Akif Khan and Shah Khusro

An Overview of OCT Techniques for Detection of OphthalmicSyndromes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109Adeel M. Syed, Muhammad Usman Akbar, and Joddat Fatima

Fully Automated Identification of Heart Sounds for the Analysisof Cardiovascular Pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117Ghafoor Sidra, Nasim Ammara, Hassan Taimur, Hassan Bilal,and Ahmed Ramsha

Modeling and Simulation of Resource-Constrained VaccinationStrategies and Epidemic Outbreaks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131Rehan Ashraf, Bushra Zafar, Sohail Jabbar, Mudassar Ahmad,and Syed Hassan Ahmed

Big Data in Healthcare: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143Muhammad Mashab Farooqi, Munam Ali Shah, Abdul Wahid,Adnan Akhunzada, Faheem Khan, Noor ul Amin, and Ihsan Ali

Internet of Things-Based Healthcare:Recent Advances and Challenges. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153Syed Tauhid Ullah Shah, Hekmat Yar, Izaz Khan, Muhammad Ikram,and Hussain Khan

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

Quality Assessment and Classificationof Heart Sounds Using PCG Signals

Qurat-ul-ain Mubarak, Muhammad Usman Akram, Arslan Shaukat,and Aneeqa Ramazan

1 Introduction

The heart is the primary organ of the cardiovascular system that acts as a pump,which supplies oxygenated blood to the body from the lungs and deoxygenatedblood to the lungs for removal of carbon dioxide [1]. The human heart is approxi-mately the size of a large fist and is located in middle of the chest, between the lungswithin thoracic cavity [2]. It has four chambers: lower left and right ventricles andupper left and right atrium [3]. The blood pressure (BP) is created by the contractionof left ventricles. The normal BP range is 120/80–140/90 for a healthy person. Themass of an adult heart is approximately 250–350 gm. The cardiac cycle is a completeheart beat that consists of a sequence of mechanical and electrical events which arerepeated with every heartbeat. It starts with systole that is a contraction of atria orventricles and ends with diastole that is a relaxation and filling of ventricles or atriawith blood. The heart rate is the frequency of cardiac cycle, and it is expressed inbeats per minute. The normal heart rate of an adult person is from 60 to 100 beatsper minute. The heart generates the electrical activity throughout the cardiac cycleas a result of which atria and ventricles contract. The vibrations are produced by theopening and closure of valves, which are audible and can indicate the condition ofheart.

The cardiovascular diseases (CVDs) are the primary cause of deaths throughoutthe world. The CVDs have caused almost 17.5 million deaths per year in the entireworld, which is almost 31–32% of total deaths [4]. This percentage is increasingdrastically day by day especially in developing countries. According to a recent

Q. Mubarak · M. U. Akram (�) · A. Shaukat · A. RamazanDepartment of Computer Engineering, National University of Science and Technology,Islamabad, Pakistane-mail: [email protected]

© Springer Nature Switzerland AG 2019F. Khan et al. (eds.), Applications of Intelligent Technologies in Healthcare,EAI/Springer Innovations in Communication and Computing,https://doi.org/10.1007/978-3-319-96139-2_1

1

http://crossmark.crossref.org/dialog/?doi=10.1007/978-3-319-96139-2_1&domain=pdf

mailto:[email protected]

https://doi.org/10.1007/978-3-319-96139-2_1

2 Q. Mubarak et al.

study in the United States, every one in a seven deaths is caused by the CVD [5].According to the National Vital Statistics System [6], heart disease is the number 1cause of deaths among the top 15 causes in the United States during 2013 resultingin almost 611,105 deaths, which is 23.5% of total deaths. The coronary heartdiseases rank highest among the top 20 diseases in Pakistan [7]. According to theWorld Health Organization, 9.87% of total deaths in Pakistan are due to heart attackwhich ranks Pakistan #63 in the world [8].

The cardiovascular disease involves the blood vessels or heart. It normallyrefers to blockage or narrowing of blood vessels causing stroke, angina pain, orheart attack. There are three types of heart diseases: electrical, circulatory, andstructural [9]. Irregular or abnormal heartbeats called arrhythmia are caused bythe problems in electrical activity of the heart. It means a heartbeat is added andskipped and the heart is beating fast or slow [10]. Arrhythmia can be either harmfulor harmless and one might not notice them. Circulatory diseases caused by thedisorders in circulatory system such as high BP, blockage of blood vessels, etc. Thestructural disease involves heart muscles/valves and congenital problems. The maincauses of CVD are high blood pressure, poor diet, smoking, obesity, diabetes, highcholesterol, genetics, age, gender, lack of exercise, and physical inactivity [11].

CVDs can be hard to diagnose. The common methods of diagnosis includeechocardiogram, electrocardiogram (ECG), X-ray, CT heart scan, MRI, cardiaccatheterization, etc. The old and traditional method of diagnosis is auscultation,which involves the examination of heart sounds using stethoscope by a cardiologist.It is an effective technique but requires proper skills [12] and extensive practice.Some common modalities to monitor heartbeats are electrocardiography (ECG),photoplethysmography (PPG), and phonocardiogram (PCG) as shown in Fig. 1.

PCG stands for phonocardiogram. It records the mechanical activity of theheart by digital stethoscope as a high fidelity plot. PCG signals are sound signalsgenerated as a result of vibrations caused by the closure of valves. Normally thehuman heart produces two sounds, lub and dub. Lub or S1 is the first sound andoccurs at the start of systole period. Its frequency is between 25 and 45 cycles persecond, and the duration is 0.14–0.15 s. S1 is best heard at the apex of the heart withdiaphragm of stethoscope. Dub or S2 is the second heart sound and occurs at thestart of diastole period during cardiac cycle. S2 has higher pitch than S1 and louder.Its duration is 0.11–0.12 s and frequency is 50 Hz. S2 is best heard with diaphragmof the stethoscope at erb’s point. Extra heart sounds are S3 and S4. They are low -frequency sounds called gallops [13]. The areas producing S3 and S4 can be locatedon either left or right side. S3 occurs at 3rd part of diastole. Its duration is almost0.1 s and frequency is 40–50 Hz. It is best heard at mitral area with stethoscope bell.S4 occurs before S1 in late diastole period and usually heard in case of hypertension.Its frequency is less than 20 cycles per second. It is heard by a stethoscope bellabove apex area and mitral valve. An example of PCG signal with annotations ofheart sounds with respect to cardiac cycle is shown in the Fig. 2.

Quality Assessment and Classification of Heart Sounds Using PCG Signals 3

Fig. 1 PCG and ECG signal with respect to cardiac cycle [14]

Time(second)

Am

plitu

de

0-2.5

-2

-1.5

-1

-0.5

0

0.5

1

1.5

0.2

systole (Sint) diastole (Dint) diastole (Dint)systole (Sint)

S1 S2

S3S4

S1 S2

S3 S4

S1

S4

S1durS2dur

S3dur S4dur

S1dur

0.4 0.6 0.8 1 1.2

cardiac cyclecardiac cycle

1.4 1.6 1.8

Fig. 2 PCG signal with annotations [15]

2 Related Work

PCG is a comparatively new metric, and correct localization and classificationof heart sounds have been a challenging task because of inconsistency of heartcycles. Many researchers have been trying to analyze PCG signals using different

4 Q. Mubarak et al.

methods like filtering, transforms-based algorithm, denoising, feature extraction,classification, etc. Major aims of research on PCG signals are to localize heat soundsand then to classify signals as normal or abnormal.

Previous techniques used for localization include time-domain filtering, downsampling, and envelope extraction using Shannon energy [16]. S1 and S2 canbe differentiated by their time duration. The classification was done by decisiontree and multilayer perceptron. In [17] the localization of heart sounds, S1 andS2 in PCG signals are done by frequency filtering, energy detection, and intervalregulation. The CIC filtering, Hilbert transform, moving average filter is applied onthe signal in sequence after which S1 and S2 peaks are detected. RAD (replace, add,and delete) algorithm is used for correction of detected peaks. The problem withtime-dependent algorithm is that they are hard to adopt to irregular heartbeats anddoesn’t provide method for noise reduction. To solve this, Xingri Quan [18] has usedGaussian regression on smoothed simplicity profile to detect S1 and S2 from heartsounds with murmurs. The murmurs are differentiated from normal componentsby Gaussian spread and weight. One cardiac cycle is taken after smoothing ofsimplicity profile, and baseline offset is removed to apply Gaussian regression. TheGaussians which correspond to murmur are removed, and S1 and S2 are detectedfrom merging adjacent Gaussians. The success rate obtained was 82% and providesa promising and robust algorithm for detection of S1 and S2 and murmur extraction.Probabilistic models have also been used for identification of S1 and S2. An [19]algorithm based upon hidden semi-Markov model (HSMM) with logistic regressionis used for the probability estimation instead of Gaussian or gamma distribution,and the former knowledge about the estimated duration of S1 and S2 states is alsoincorporated. Extended Viterbi algorithm is also used. This algorithm successfullysegments noisy, real-world normal and abnormal pcg signal and leaves behind thecurrent state-of-the-art method based on a two-sided, paired t-test.

Presently, different machine learning algorithms such as artificial neural network,DNN, and SVD have been used for the classification. In [20] artificial neuralnetwork-based method for classification of S1 and S2, which has been localizedby optimized s-transform. SVD is used for feature extraction on the basis of whichclassification is done using back propagation algorithm. Grzegorczk [21] has alsoused the neural networks with hidden Markov’s model and achieved overall scoreof 0.79. He also successfully coped with artifacts and eliminated the effects ofdisturbances and interruptions. Reliable S1 and S2 segmentation, in cases wherethe time interval of S1 and S2 is unknown, can be done by using only acousticcharacteristics [22]. The recognition is based upon deep neural networks. Thealgorithms using neural networks show better results than conventional methodsand also reduce energy spreading problem.

Despite the advancement of research on segmentation of PCG, the precise andcorrect classification of S1 and S2 is still a challenging task, and research gap isavailable regarding the application of machine learning in this field. The multi-domain analysis can be done. Both time-domain and frequency domain featurescan be used to improve segmentation accuracy. Moreover the signal quality canbe assessed before using them for processing as the PCG signals are likely to bedistorted and noisy during accusation.


3 Dataset

Different PCG signal datasets are available publicly to encourage especiallymachine learning researchers to work on pcg signals. PhysioNet [23] and Pascal [24]have arranged the Classification of Normal/Abnormal Heart Sound Recordings: thePhysioNet/Computing in Cardiology Challenge 2016 and Classifying Heart SoundChallenge 2012, respectively. The basic aim of these challenges was to classifyPCG signals into normal and abnormal classes using segmentation of heart sounds.The detail of the datasets provided is explained in Table 1.

4 Methodology

In this section, we will describe the methodology adopted for the correct localizationand classification of S1 and S2. The dataset from Pascal classifying heart soundchallenge is used. We have included the quality assessment step prior to processing

Table 1 Specifications of dataset provided by Pascal and PhysioNet challenge

Pascal PhysioNetAcquisition Dataset A Dataset B Collected at either a clinical ormethod From the general public From a clinic trial in nonclinical environment, from

via the iStethoscope Pro hospitals using the digital both healthy subjects and

iPhone app stethoscope digi scope pathological patients

Uncontrolled environmentDataset Folders No. of

samplesFolders No. of

samplesFolders No. of

samplesNormal 31 Normal 320 Training set:

5 folders (A – E)3126

Murmur 34 Murmur 95Extra heart 19 Extrasystole 46Sound Validation set 300Artifact 40 B unlabeled

test195

A unlabeledtest

52

Annotations Normal (1) Abnormal (−1)Unsure/too noisy

Challenge 1 S1 and S2 S1 and S2Challenge 2 Normal

MurmurExtra heartsoundArtifact

NormalMurmurExtrasystole

Resolution 44,100 Hz 4000 Hz 2000 Hz

6 Q. Mubarak et al.

Fig. 3 Flow diagram of proposed methodology

in order to evaluate the fitness of the signal for analysis. Next we have segmented theheart sound locations from the PCG signals and extracted features for them whichare fed to classifier to classify heart sounds between S1 and S2. Figure 3 shows theflow of proposed method.

4.1 Quality Assessment

Before applying any processing on the signals, the quality of the signals is assessedto determine the suitability of signal according to predefined criteria. For evaluation,first of all wavelet coefficients of signal is calculated by discrete wavelet transform


using Daubechies-2 at 2nd decomposition level. The approximation coefficients ofthe 2nd level are used for evaluation. The signal is evaluated on the basis of threecriteria.

1. The successive difference of the signal is calculated, and the root mean square istaken of the obtained difference, which should be equal or less than one.

2. The no. of zero crossings, .i.e., the number of intersection of signals with the x-axis, is determined. 0.85 quantile of the no. of zero crossing divided by the totallength of the signal should be less than 0.05.

3. The signal is divided into windows of length 2205 ms, and if no. of peaks in awindow is less than 15, the window is assigned to 1. The signal falls on criteria,if the signal has 65% of windows with 1.

The signal is considered as suitable, if it falls on all three criteria otherwisereacquisition is required for signals, which do not fulfill the criteria.

4.2 Localization

The location of heart sounds is manually extracted from annotations given by thePascal along with database. Five thousand samples were taken from the front andback of the annotations each (10,000 samples in total) and stored in the matrix. Eachrow contains the samples of heart sounds. The locations were obtained from eachfile of Atraining_normal files of dataset provided by Pascal.

4.3 Feature Extraction

Both time and frequency domain features have been used for the analysis. Thefeatures for each heart sound are calculated, and labels have been assigned to them.Features of S1 heart sounds are assigned 0 and features of S2 is assigned 1. Thefollowing features have been used:

1. S1 std/total std: Ratio of the standard deviation of S1 over the total standarddeviation of whole signal [16]

2. S2 std/total std: Ratio of the standard deviation of S2 over the total standarddeviation of whole signal [16]

3. S1 mean/total mean: Ratio of the mean of S1 over total mean of whole signal[16]

4. S2 mean/total mean: Ratio of the mean of S2 over total mean of whole signal[16]

5. S1 FFT/total FFT: Ratio of fast Fourier transform of S1/total fast Fouriertransform of signal [16]

8 Q. Mubarak et al.

6. S2 FFT/total FFT: Ratio of fast Fourier transform of S2/total fast Fouriertransform of signal [16]

7. Kurtosis S1: Measures the sharpness of peaks in S18. Kurtosis S2: Measures the sharpness of peaks in S29. Shannon entropy of approx1_S1: Shannon entropy of the approximation coeffi-

cient obtained by performing DWT on S1 using Daubechies-2 at 1st level10. Shannon entropy of detail1_S1: Shannon entropy of the detail coefficients

obtained by performing DWT on S1 using Daubechies-2 at 1st-level Shannonentropy of the approximation coefficient obtained by performing DWT on S1using Daubechies-2 at 2nd level

11. Shannon entropy of approx2_S1: Shannon entropy of the approximation coef-ficient obtained by performing DWT on S1 using Daubechies-2 at 2nd level

12. Shannon energy of approx1_S1: Shannon energy of the approximation coeffi-cient obtained by performing DWT on S1 using Daubechies-2 at 1st level

13. Shannon energy of detail1_S1: Shannon energy of the detail coefficientsobtained by performing DWT on S1 using Daubechies-2 at 1st level

14. Shannon energy of approx2_S1: Shannon energy of the approximation coeffi-cient obtained by performing DWT on S1 using Daubechies-2 at 2nd level

15. Shannon entropy of approx1_S2: Shannon entropy of the approximation coeffi-cient obtained by performing DWT on S2 using Daubechies-2 at 1st level

16. Shannon entropy of detail1_S2: Shannon entropy of the detail coefficientsobtained by performing DWT on S2 using Daubechies-2 at 1st level

17. Shannon entropy of approx2_S2: Shannon entropy of the approximation coef-ficient obtained by performing DWT on S2 using Daubechies-2 at 2nd level

18. Shannon energy of approx1_S2: Shannon energy of the approximation coeffi-cient obtained by performing DWT on S2 using Daubechies-2 at 1st level

19. Shannon energy of detail1_S2: Shannon energy of the detail coefficientsobtained by performing DWT on S2 using Daubechies-2 at 1st level

20. Shannon energy of approx2_S2: Shannon energy of the approximation coeffi-cient obtained by performing DWT on S2 using Daubechies-2 at 2nd level

4.4 Classification

KNN classifier is used for classification of heart sounds into S1 and S2. The featuresof all files and both S1 and S2 are concatenated, and labels are extracted, and bothdata and labels are shuffled. The fivefold cross validation is applied. The data isdivided into fivefolds, and KNN classifier is applied on each fold. The k is variedfrom 1:5, and accuracy for each k is calculated for each fold.


Table 2 Averaged performance evaluation parameters

Evaluation parameter Before quality assessment After quality assessment

Accuracy 0.86 ± 0.0014 0.88 ± 0.00117Sensitivity 0.80 ± 0.0002 0.83 ± 0.007Specificity 0.89 ± 0.006 0.92 ± 0.0.009Precision 0.89 ± 0.007 0.92 ± 0.007F1 score 0.83 ± 0.008 0.87 ± 0.005

Fig. 4 Confusion matrix (a) before quality assessment and (b) after quality assessment

5 Results

The accuracy sensitivity, specificity, precision, F1 score, and confusion matrix areused for the evaluating performance of classifier. The KNN classifier is applied firstwithout quality assessment of signal and then applying the quality assessment andleaving the non-suitable signal. A total 27 signals out of 31 were marked suitable.The average accuracy is improved from 0.86 ± 0.0014 to 0.88 ± 0.00117. Theperformance parameters before and after quality assessment are shown in Table 2.The confusion matrix is shown in Fig. 4.

6 Conclusion

The algorithm provides a suitable criterion for the quality assessment for PCGsignals before any processing. The results show that performance is increased byleaving the non-suitable/too noisy signals before applying any processing. Theresults can further be improved by integrating both time- and frequency-domainfeatures and using machine learning state-of-the-art techniques like convolutionalneural network, logistic regression, particle swarm optimization, etc. The effectsof different decompositions like wavelet decomposition (WD), time-frequencydecomposition (TFD), and empirical mode decomposition (EMD) can be studiedand applied.

Acknowledgments The author would like to acknowledge Biometrics, Medical Image, and SignalAnalysis (BIOMISA) research group and Dr. Ahsan Imran and Dr. Hina Ayub for providingassistance and facilitating the research process.

10 Q. Mubarak et al.

References

1. Hoffman, M. Human anatomy. http://www.webmd.com/heart/picture-of-the-heart2. Heart anatomy. https://cnx.org/contents/Y5T_wVSC@3/Heart-Anatomy3. Sherwood, L. (2012). An Physiology: from cells to systems, 8, revised ed., Cengage Learning,

p. 928.4. World Health Organization. (2016). Cardiovascular diseases (CVDs): fact sheet. http://

www.who.int/mediacentre/factsheets/fs317/en/5. Mozaffarian, D., et al. (2015). Heart disease and stroke statistics—2015 update: a report

from the American heart association. Circulation. 131(4):e29–e322. https://www.heart.org/idc/groups/ahamah-public/@wcm/@sop/@smd/documents/downloadable/ucm_470704.pdf

6. National Vital Statistics System. https://www.cdc.gov/nchs/nvss/7. Top 20 Causes Of Death Pakistan. (2014). http://www.worldlifeexpectancy.com/pakistan-

coronary-heart-disease8. Nishtar, D. S. (2001). The CVD situation in Pakistan. http://www.heartfile.org/pdf/

Essentialdrugs.pdf9. Heart Diseases & Disorders. (2017). http://www.hrsonline.org/Patient-Resources/Heart-

Diseases-Disorders?gclid=CPHnkoOXz8sCFRG6Gwod8WAC7A10. Abnormal Heart Rhythms (Arrhytmias). http://www.webmd.com/heart-disease/guide/heart-

disease-abnormal-heart-rhythm#111. Division for Heart Disease and Stroke Prevention: Centre for Disease control and prevention.

https://www.cdc.gov/dhdsp/data_statistics/fact_sheets/fs_heart_disease.htm12. Hauora, T. P. (2013). CDHB clinical skills unit – heart and lung auscultation.13. Basic Heart Sounds. http://www.stethographics.com/main/physiology_hs_introduction.html14. Best, B. Electrocardiogram (ECG, EKG) Interpretation emphazing ST-segment and T-wave

deviation (ischemic changes). https://www.benbest.com/health/ECG.html#ECG15. Varghees, N., & Ramachandran, K. I. (2014). A novel heart sound activity detection framework

for automated heart sound analysis. Biomedical Signal Processing and Control, 13, 174–188.16. Gomes, E. F., Jorge, A. M., & Azevedo, P. J. (2013). Classifying heart sounds using

multiresolution time series motifs: An exploratory study. Portugal: Porto.17. Dong S. M., & Shin H. (2015). A localization method for first and second heart sounds based

on energy detection and interval regulation. Journal of Electrical Engineering and Technology,10(5), no. ISSN (Print) 1975-0102, pp. 2126–2134.

18. Quan, X., Seok, J., & Bae, K. (2015). Detection of S1/S2 components with extraction ofmurmurs from phonocardiogram. IEICE Transactions on Information & Systems, 98-D(3),745–748.

19. Springer, D. B., Tarassenko, L., & Clifford, D. G. (2016). Logistic regression-HSMM-basedheart sound segmentation. IEEE Transactions on Biomedical Engineering, 63(4), 822–832.

20. Shivhare, V. K., Sharma, S. N., & Shakya, D. K. (2015). Detection of heart sounds S1 andS2 using optimized S-transform and back-propagation algorithm. In: IEEE Bombay SectionSymposium (IBSS), Mumbai.

21. Grzegorczk, I., Solinski, M., Łepek, M., Rymko, J., Rymko, J., Stepien, K., & Gierałtowski,J. (2016). PCG classification using a neural network approach. In: Computing in cardiology,Vancouver, Canada.

22. Chen, T., Yang, S., Ho, L., Tsai, K., Wang, S., Chen, Y., Lai, Y., Chang, Y., & Wu, C. (2016).S1 and S2 heart sound recognition using deep neural networks. In: IEEE Transactions onBiomedical Engineering.

23. PhysioNet. Classification of Normal/Abnormal Heart Sound Recordings: the Phys-ioNet/Computing in Cardiology Challenge 2016. https://physionet.org/challenge/2016/

24. Bentley, P. (2012). Pascal Classifying Heart Sounds Challenge. http://www.peterjbentley.com/heartchallenge/

http://www.webmd.com/heart/picture-of-the-heart

https://cnx.org/contents/Y5T_wVSC@3/Heart-Anatomy

http://www.who.int/mediacentre/factsheets/fs317/en/

https://www.heart.org/idc/groups/ahamah-public/@wcm/@sop/@smd/documents/downloadable/ucm_470704.pdf

https://www.cdc.gov/nchs/nvss/

http://www.worldlifeexpectancy.com/pakistan-coronary-heart-disease.(2014)

http://www.heartfile.org/pdf/Essentialdrugs.pdf

http://www.hrsonline.org/Patient-Resources/Heart-Diseases-Disorders?gclid=CPHnkoOXz8sCFRG6Gwod8WAC7A

http://www.webmd.com/heart-disease/guide/heart-disease-abnormal-heart-rhythm#1

https://www.cdc.gov/dhdsp/data_statistics/fact_sheets/fs_heart_disease.htm

http://www.stethographics.com/main/physiology_hs_introduction.html

https://www.benbest.com/health/ECG.html#ECG

https://physionet.org/challenge/2016

http://www.peterjbentley.com/heartchallenge/


Qurat-ul-ain Mubarak received her BS degree in Electronic Engineering from the InternationalIslamic University, Islamabad, in 2014. She is currently enrolled in MS program in the field ofComputer Engineering at C of EME, NUST, Pakistan. Her area of interest is Image Processing andsignal processing specifically biomedical signals.

Muhammad Usman Akram is an assistant professor at C of EME, NUST, Pakistan. He hasPhD degree in computer engineering with specialization in medical image processing. His areas ofinterest are image processing, signal processing, and machine learning.

Arslan Shaukat is an assistant professor at C of EME, NUST, Pakistan. He has PhD degree fromthe United Kingdom (UK) with specialization in machine learning. His areas of interest are imageprocessing, signal processing, and machine learning.

Aneeqa Ramazan received her BS degree in Electronic Engineering from the InternationalIslamic University, Islamabad, in 2015. She is currently enrolled in MS program in the fieldof Computer Engineering at CEME, NUST, Pakistan. Her area of interest is Image Processingspecifically biomedical signals and signal processing.

Classification of Normal Heart BeatsUsing Spectral and Nonspectral Featuresfor Phonocardiography Signals

Shahid Ismail Malik and Imran Siddiqi

1 Introduction

PCG or phonocardiography is the pictorial representation of acoustic activity ofthe heart. Because of the repetitive nature of heart cycle, it can be segmentedinto four major parts in case of a healthy person having normal heart beat. Thesefour segments are the first heart beat (S1), systole, second heart beat (S2), anddiastole. In case of an abnormality, murmurs (noisy sounds) are also introducedin the normal cycles, or some normal beat is skipped. Each of these abnormalitiesrefers to a particular problem in the heart. Analysis of these signals serves as aneffective tool for the diagnosis of a number of heart diseases or for initiating furtherinvestigation. Localization and classification of these sounds has been studied fordecades. The problem, however, still remains challenging due to the complexityit offers. An automated system for analysis of heart recordings relies on twomain components, classification and localization. Localization involves locating thespecific heart sounds in the signal, while classification assigns a label to each of thelocalized sounds. As a function of the methodology, classification can be followedby localization or vice versa. An example PCG signal is shown in Fig. 1.

Heart beat analysis has been an active area of research for many years.Researchers focus either on classification or localization and classificationsimultaneously. Among well-known contributions, Abbas and Bassam [1] employed

S. Ismail MalikArmy Public College of Management and Sciences, Rawalpindi, Pakistan

Bahria University, Islamabad, Pakistan

I. Siddiqi (�)Bahria University, Islamabad, Pakistane-mail: [email protected]


13



https://doi.org/10.1007/978-3-319-96139-2_2

14 S. Ismail Malik and I. Siddiqi

Fig. 1 Phonocardiographic or PCG of human heart [6]

the time-frequency approach for classification and localization of heart sounds. Theinvestigated features include peaks of short-time Fourier transform (STFT) andintensity and realized promising results. Safara et al. [12] first preprocess the signaland extract features using wavelet packet transform. Principal component analysis(PCA) is then applied, and the transformed feature set is fed to a hybrid classifiercomprising support vector machine (SVM), k-nearest neighbor (KNN), and threemultilayer perceptrons (MLPs) to predict heart valve disorders (HVD). The workwas later extended to investigate the wavelet packet entropy for heart murmurclassification [11].

In another study, Amir and Armano [3] first preprocessed the signal using low-pass filtering and decimation followed by peak conditioning using complex Morletwavelet (CMW). Maximum value amplitude, Shannon energy, bispectrum, andWigner bispectrum features were then extracted from resultant signals. The signalswere finally classified as normal or abnormal using classification and regressiontrees (CART) algorithm. An adaptive fuzzy inference system is proposed in [2]where a Mamdani-type fuzzy classifier was developed to detect the presence ofmurmurs. Likewise, Wu et al. [18] employed Mel frequency cepstral coefficients(MFCC)-based hidden Markov models (HMMs) for the identification of heartdiseases. In another significant work, Springer et al. [14] recorded heart signals from3M Littmann 3200 electronic stethoscope and an iPhone 3G. The signal quality wasgreatly explored and subsequently classified using SVM.

A number of studies [4, 5, 7, 9, 10, 16, 17] borrowed the segmentation algorithmproposed by Springer et al. [15] and focused on the classification part only. TheSpringer’s segmentation algorithm is based on the ECG signal. The R and T wavesin ECG correspond to S1 and S2 in the PCG. This correspondence was exploitedin [9] along with hidden semi-Markov models (HSMM) and extended Viterbi algo-rithm. HSMM was used for modeling the expected duration in which the heart staysin S1, S2, systole, and diastole states, while the Viterbi algorithm was employedfor decoding the most likely sequence of these states. In a similar study, Banerjeeet al. [5] studied the Mel frequency cepstral coefficients (MFCCs) and waveletfeatures for classification, while features from time, frequency spectrum, energy,cyclostationarity, and power spectral density are investigated in [16]. Among other

Classification of Normal Heart Beats Using Spectral and Nonspectral Features. . . 15

studies, Whitaker and Anderson [17] investigated sparse coding for classification,while Antik et al. [4] employed nonnegative matrix factorization (NMF) and randomforests for classification. Runnan et al. [8] used AR model of A2 of S2 beat todiscriminate between normal and abnormal heart sounds, while Puri et al. [10] usedstatistical, wavelet packet energy and spectral features for classification. In anotherwell-known study, Grzegorczyk [7] investigated time, frequency, and ordinate axis(morphological) features for classification and reported promising classificationrates. This paper presents an effective technique for classification of PCG signalsusing spectral and nonspectral features. Features based on spectrum, energy, andprobability are considered in our study, while classification is carried out using afeed-forward artificial neural network. The details of the proposed technique alongwith the dataset are presented in Sect. 2, while Sect. 3 discusses the experiments andthe realized results. Finally, conclusions are drawn in Sect. 4 with a discussion onfuture research directions on this problem.

2 Methodology

This section presents the details of the proposed methodology for classification ofheart beats from PCG signals. An overview of the proposed technique is illustratedin Fig. 2. The system reads data files along with the actual beat location, and thepeak value from the ground truth is centralized in a window of width 2 × 3807making a total of (2 × 3807 + 1) 7615 samples. The process is presented in Fig. 3.Features are extracted by generating a Gaussian window around the peak value.Principal component analysis (PCA) is then applied on the extracted features, andthe transformed features are fed to an artificial neural network for training andsubsequently for classification. Each of these steps is discussed in the followingsections.

2.1 Dataset

The dataset used in our study is a subset of the PASCAL Classifying Heart SoundsChallenge (CHSC) 2011 database [6]. The dataset was collected by iStethoscope ProiPhone app and was labeled as Dataset-A by the developers. Evaluations carried out200 normal heart beat signals from the dataset with 100 S1, 100 S2, and 100 noise

Read Data andGround Truth Files

Centralize Peaks &Calculate Features

DimensionalityReduction

ANN basedClassification

Fig. 2 Overview of the proposed system


0-0.2

0y[n]

0.2

1000 2000 3000 4000x[n] Samples

S1 Beat

5000 6000 7000

Hamming Window

Rectangular Window

Non Centered Rectangular Window

8000

0-0.2

0y[n]

0.2

1000 2000 3000 4000 5000 6000 7000 8000

0-0.2

0y[n]

0.2

1000 2000 3000 4000 5000 6000 7000 8000

Fig. 3 Peak value centering in a window

labels making 300 samples. The labels and locations of S1 and S2 are provided inthe ground truth files.


Features considered in our study can be divided into three broad categories. Theseinclude:

– FFT and spectral features– Energy-based features– Statistical and probability features

A brief discussion on these features is presented in the following.

FFT and Spectral Features

Fast Fourier transform (FFT) is applied on the Gaussian window of 7615 samples.The sampling frequency (Fs) is 44100 giving us frequency components from DCto 22050 Hz. Out of these components, we keep 180 points of the FFT from 20 to


199 Hz. We also divide the spectral components of these 180 FFT points furtherinto 7 spectral bands. The exact range of each band can be found in Table 1.Spectral correlation between these 7 bands is also computed producing 21 additionalfeatures. Hence the total spectral features are 381 (180 FFT points, 180 points FFTbased 7 spectral bands and 21 correlations features) as shown in Table 3.

Probability-Based Features

CDF number (CDF#) is a novel feature introduced in this study and is calculated byfirst normalizing the signal to four decimal points. As discussed earlier, the signallength is 7615. The probability density function (PDF) of the signal is first calculated(Fig. 4a). CDF# is the total number of distinct values in the PDF. For instance, forthe S1 beat in Fig. 4a, this number is 1064. Figure 4a–c shows that the values in theS1 beat are roughly in the range from −0.12 to 0.078. Similarly the PDFs for S2and noise are shown in Fig. 4b, and the corresponding CDF# can be calculated. Theother probability-based features include signal (beat or noise) mean and standarddeviation and mean and standard deviation of the PDF. This gives a total of five

Table 1 An overview of the features employed in our study

SNo. Domain Dimensionality Description

1 Spectrum FFT = 180 180 point FFT of signal from 20–199 Hz

SPx = 180 where x =[1. . . 7]

sp1 = [20–49 Hz], sp2 = [50–74 Hz]

SP1 = 30, [SP2. . . SP7] =25

sp3 = [75–99 Hz], sp4 = [100–124 Hz]

Spectral band correlation= 21

sp5 = [125–149 Hz], sp6 = [150–174 Hz]

sp7 = [176–199 Hz]

Spectral band correlation: correlationbetween spx’s

2 Probability Mean = 1 Mean of signal in Gaussian window

Standard deviation = 1 Standard deviation of signal in gaussian win-dow

CDF# = 1 CDF number

Mean(PDF) = 1 Mean of PDF window for CDF#

Standard deviation(PDF) = 1

Standard deviation of PDF window for CDF#

3 Energy Power = 1 Power of signal from 24–199 Hz

Normalized power = 1 Normalized power of signal from 24–199 Hz

Power percentage = 1 90% power of signal from 24–199 Hz

Power length = 1 Length for 90% power of signal from 24–199 Hz

Product of power andlength = 1

Product of 90% power & length of 90%power


Fig. 4 CDF# calculation and example PDF plot for S1, S2, and noise. (a) Calculation of CDFnumber. (b) PDFs of S1, S2, and noise

Table 2 Values of CDF features for S1, S2, and noise

SNo. S1 S2 Noise

CDF# 1064 704 301

Mean(PDF) 7.15 10.81 25.29

Standard deviation (PDF) 16.00 21.53 36.11

Mean (Signal) −3.755e–04 −3.217e–05 −7.162e–05

Standard deviation (Signal) 0.0198 0.0125 0.0042

probability-based features. Values of these features for the samples in Fig. 4b aresummarized in Table 2.

Energy-Based Features

Five energy-based features are considered in our study. These include total power,power of normalized beat or noise from 24 to 199 Hz, 90% power of signal, lengthof 90% power, and product of 90% power and length of 90% power.

A summary of the features employed in our study is presented in Table 1.

2.3 Dimensionality Reduction

Dimensionality reduction is generally applied to reduce the feature size whilekeeping only useful features. Dimensionality reduction includes feature selection(selecting a subset of useful features) and feature extraction (transforming thefeatures to a new space of reduced dimensionality). Among well-known featureextraction techniques, principal component analysis (PCA), linear discriminant


analysis (LDA), etc. have been widely employed. In our study, we have chosen toapply PCA [13] for dimensionality reduction by preserving 95% of variance in thedata. The classification rates with and without PCA are discussed in Sect. 3.

2.4 Classification

For classification, we have chosen to employ a feed-forward artificial neural network(ANN). Seventy percent of data is used for training, while the remaining 30% is usedfor testing. The performance of the system is studied as a function of the number ofhidden neurons for individual features as well as the combination of features.

3 Results

This section presents the results of the evaluations carried out to study theeffectiveness of the proposed features. We first present the performance (in terms ofclassification rate) of the individual as well as combined features with and withoutPCA as a function of the neurons in the hidden layer. We also present the respectiveconfusion matrices, sensitivity, specificity, and precision of different classes. Theclassification rates using probability and energy-based features are illustrated inFig. 5, while those based on spectrum-based features and the combination of allfeatures are presented in Fig. 6.

A summary of the realized classification rates with and without PCA is presentedin Table 3. It can be seen that the spectral features outperform the energy- andprobability-based features reporting classification rates of 86.59% and 80.12% withand without PCA, respectively. Using the combination of all features, a maximumclassification rate of up to 92.33% is realized. It is also interesting to note that byreducing the dimensionality of the combined features from 391 to 36 (more than90% reduction), the system reports a maximum classification rate of 89.33% whichindeed is very promising.

Figures 7 and 8 present the summarized results in terms of confusion matrices,sensitivity, specificity, and precision of different classes using individual andcombined features with and without application of PCA. It can be seen that theperformance of the probability-based features is the least impressive, while thespectral features report the best results among individual features. The combinationof all features naturally outperforms the individual features. In general, specificityof all classes is well above 0.8, while the sensitivity is 0.75 or more for all classesexcept for S2 where it drops to 0.52 for probability-based features.


Tab

le3

Sum

mar

yof

clas

sific

atio

nra

tes

wit

han

dw

itho

utPC

A

Feat

ures

wit

hout

PCA

Feat

ures

afte

rPC

Aap

plic

atio

n

% Dim

ensi

onal

ity

redu

ctio

n

Ave

rage

clas

si-

ficat

ion

wit

hout

PCA

Max

imum

clas

sific

atio

nw

itho

utPC

A

Ave

rage

clas

sific

atio

nw

ith

PCA

Max

imum

clas

sific

atio

nw

ith

PCA

Feat

ure

dom

ain

Ene

rgy

55

075

.28%

81.6

71%

75.8

0%81

.67%

Prob

abil

ity

53

4063

.75%

72.6

7%63

.59%

69%

Spec

tral

381

3590

.81

86.5

9%91

%80

.12%

88.6

7%

All

feat

ures

391

3690

.79

84.1

9%92

.33%

81.3

2%89

.33%


Fig. 5 Classification based on energy and probability. (a) Energy-based features with PCA. (b)Energy-based features without PCA. (c) Probability-based features with PCA. (d) Probability-based features without PCA

4 Conclusion

This study investigated the effectiveness of spectral-, energy-, and probability-based features to classify heart beat sounds into S1, S2, and noise. Classificationis carried out using an artificial neural network with and without applying PCAto the extracted feature set. Evaluations carried out on the publicly availablePASCAL-CHSC database reveal that the spectral features outperform energy- andprobability-based features. It was also observed that high classification rates aremaintained by using only a small proportion of the feature set. The present studyfocused on the classification of beats only, and the location of the beats was taken


Fig. 6 Classification based on spectrum-based features and combination of all features. (a)Spectrum-based features with PCA. (b) Spectrum-based features without PCA. (c) All featureswith PCA. (d) All features without PCA

from the ground-truth data. In our further study, we intend to research localizationtechniques which can be combined with the classification module. We also plan toevaluate the system on the complete database and compare our results with state-of-the-art localization and classification techniques.


S1 S2 Noise S1 S2 Noise S1 S2 Noise S1 S2 NoiseS1 91 4 5 90 4 6 94 0 6 94 1 5S2 21 52 27 9 75 16 1 90 9 1 94 5Noise 17 8 75 11 9 80 4 7 89 3 8 89

(a)

Sen. Spe. Prec. Sen. Spe. Prec. Sen. Spe. Prec. Sen. Spe. Prec.S1 0.91 0.947 0.705 0.9 0.947 0.81 0.94 0.97 0.949 0.94 0.97 0.959S2 0.52 0.797 0.812 0.75 0.949 0.85 0.9 0.95 0.927 0.94 0.969 0.912Noise 0.75 0.87 0.7 0.8 0.898 0.78 0.89 0.943 0.855 0.89 0.945 0.898

(b)

Probability features Energy features Spectral features All features

Probability features Energy features Spectral features All featuresO

utpu

t Lab

el Target Label Target Label Target Label Target Label

Fig. 7 Performance of individual and combined features on different classes without PCA. (a)Confusion Table of different feature sets with out using PCA. (b) Sensitivity (Sen.), Speci-ficity(Spe.) and Precision (Pre.) for different feature sets without using PCA

S1 S2 Noise S1 S2 Noise S1 S2 Noise S1 S2 NoiseS1 88 9 3 95 2 3 88 5 7 95 2 3S2 7 70 23 12 70 18 3 90 7 2 93 5Noise 18 12 70 15 9 76 3 6 91 11 8 81

(a)

Sen. Spe. Prec. Sen. Spe. Prec. Sen. Spe. Prec. Sen. Spe. Prec.0.88 0.935 0.778 0.95 0.97 0.78 0.88 0.94 0.936 0.95 0.973 0.880.7 0.856 0.769 0.7 0.863 0.86 0.9 0.949 0.891 0.93 0.964 0.90.7 0.852 0.729 0.76 0.881 0.78 0.91 0.953 0.867 0.81 0.9099 0.91

(b)

Probability features Energy features Spectral features All features

Probability features Energy features Spectral features All featuresTarget Label Target Label Target Label Target Label

Fig. 8 Performance of individual and combined features on different classes using PCA (a)Confusion Table of different feature sets using PCA. (b) Sensitivity (Sen.), Specificity(Spe.) andPrecision (Pre.) for different feature sets using PCA

References

1. Abbas, A. K., & Bassam, R. (2008). PCG spectral pattern classification: Approach to cardiacenergy signature identification. In Proceedings of the 13th International Conference onBiomedical Engineering (TBME).

2. Ahmad, T. J., Ali, H., & Khan, S. A. (2009). Classification of phonocardiogram using anadaptive fuzzy inference system. In Proceedings of the International Conference on ImageProcessing, Computer Vision & Pattern Recognition (IPCV).

3. Amiri, A. M., & Armano, G. (2013). Detection and diagnosis of heart defects in newbornsusing CART. Journal of Life Sciences and Technologies, 1(3), 103–106.


4. Antink, C. H., Becker, J., Leonhardt, S., & Walter, M. (2016). Nonnegative matrix factorizationand random forest for classification of heart sound recordings in the spectral domain. InProceedings of the Computing in Cardiology Conference (CinC).

5. Banerjee, R., Biswas, S., Banerjee, S., Choudhury, A. D., Chattopadhyay, T., Pal, A.,Deshpande, P., & Mandana, K. M. (2016). Time-frequency analysis of phonocardiogram forclassifying heart disease. In Proceedings of the Computing in Cardiology Conference (CinC).

6. Bentley, P., Nordehn, G., Coimbra, M., Mannor, S., & Getz, R. (2011). Classifying heart soundschallenge, CHSC. http://www.peterjbentley.com/heartchallenge/.

7. Grzegorczyk, I. (2016). PCG classification using a neural network approach. In Proceedings ofthe Computing in Cardiology Conference (CinC).

8. He, R., Zhang, H., Wang, K., Li, Q., Sheng, Z., & Zhao, N. (2016). Classification of heartsound signals based on AR model. In Proceedings of the Computing in Cardiology Conference(CinC).

9. Munia, T. T. K., Tavakolian, K., Verma, A. K., Zakeri, V., Khosrow-Khavar, F., Fazel-Rezai,R., & Akhbardeh, A. (2016). Heart sound classification from wavelet decomposed signalusing morphological and statistical features. In Proceedings of the Computing in CardiologyConference (CinC).

10. Puri, C., Ukil, A., Bandyopadhyay, S., Singh, R., Pal, A., Mukherjee, A., & Mukherjee, D.(2016). Proceedings of classification of normal and abnormal heart sound recordings throughrobust feature selection. In Proceedings of the Computing in Cardiology Conference (CinC).

11. Safara, F., Doraisamy, S., Azman, A., Jantan, A., & Ranga, S. (2012). Wavelet packet entropyfor heart murmurs classification. Journal of Advances in Bioinformatics, 2012, 1–6.

12. Safara, F., Doraisamy, S., Azman, A., Jantan, A., & Ranga, S. (2013). Diagnosis of heartvalve disorders through trapezoidal features and hybrid classifier. International Journal ofBioscience, Biochemistry and Bioinformatics (IJBBB), 3(6), 662–665.

13. Sonka, M., Hlvac, V., & Boyle, R. (2008). Digital image processing and computer vision.Delhi: Cengage Learning India Pvt Ltd.

14. Springer, D. B., Brennan, T., Zhlke, L. J., Abdelrahman, H. Y., Ntusi, N., Clifford, G. D.,Mayosi, B. M., & Tarassenko, L. (2014). Signal quality classification of mobile phone-recordedphonocardiogram signals. In Proceedings of the IEEE International Conference on Acoustics,Speech and Signal Processing (ICASSP).

15. Springer, D. B., Tarassenko, L., & Clifford, G. D. (2016). Logistic regression-HSMM-based heart sound segmentation. In Proceedings of the IEEE Transactions on BiomedicalEngineering (TMBE).

16. Tang, H., Chen, H., Li, T., & Zhong, M. (2016). Classification of normal/abnormal heartsound recordings based on multi-domain features and back propagation neural network. InProceedings of the Computing in Cardiology Conference (CinC).

17. Whitaker, B. M., & Anderson, D. V. (2016). Heart sound classification via sparse coding. InProceedings of the Computing in Cardiology Conference (CinC).

18. Wu, H., Kim, S., & Bae, K. (2010). Hidden Markov model with heart sound signals foridentification of heart diseases. In Proceedings of the 20th International Congress on Acoustics(ICA).


Segmentation of Chest Radiographs forTuberculosis Screening Using KernelMapping and Graph Cuts

Ayesha Fatima, Anam Tariq, Mahmood Akhtar, and Hira Zahid

1 Introduction

Segmentation plays an important role in biomedical imaging to extract relevantinformation from an image. Detection of lung region from chest radiographs is animportant task in computer-aided diagnosis for the detection of different diseases,e.g., tuberculosis and lung cancer. To detect different diseases in chest radiographs,it is important to detect lung boundaries accurately. Interpretation of radiographshas become an important challenge due to its anatomical structure. Sometimes, evenexperienced radiologists have difficulty in finding severity rate in abnormal struc-tures. Due to the clinical importance of chest radiographs, it gave a new directionto research for the development of an algorithm to assist radiologists in readingradiographs [7]. Many lung segmentation techniques have been proposed. Segmen-tation techniques depend upon the characteristics of problems being considered.Otsu threshold algorithm was proposed in which an image was segmented by local,global, or adaptive threshold [13]. Pixels having values greater than threshold belongto the foreground, and the rest of the pixels belongs to the background. Varioustechniques of lung segmentation, i.e., active shape models, pixel classification, andvarious combinations thereof, were compared in [18, 20], and pixel classificationprovided efficient result on the test data. Another iterative segmentation approachwas proposed, based on intensity information and shape prior on the test data [5]. In[19], lung was subdivided into different regions of different sizes. Multi-scale filterbank was used to detect abnormal signs of texture. Different sets were constructedfor each region and classified on the basis of voting and weighted integration.

A. Fatima (�) · A. Tariq · M. Akhtar · H. ZahidDepartment of Computer Engineering, College of Electrical and Mechanical Engineering, NUST,Islamabad, Pakistane-mail: [email protected]


25



https://doi.org/10.1007/978-3-319-96139-2_3

26 A. Fatima et al.

Different supervised methods, i.e., watershed segmentation, active shape model(ASM), and graph cut method, were also discussed in [13].

Some methods used binary information, and some used regional information,but graph cut utilized both binary and regional information. Another algorithm wasproposed in [3] by Cootes, in which contour was presented as a polygon withfixed labeled points. An ASM (active shape model) is a collection of landmarkpoints and gray levels. Both the gray-level models and shape models were trainedwith the set of labeled images. In the improvement made in ASM (active shapemodel) in [21], shape model was extracted from body location using thin splinemethod, and the Gaussian pyramid was used for quick iteration. Another techniquenamed 4D lung segmentation, which is an extended form of ASM (active shapemodel), was proposed in which 4D lung segmentation is refined by optimalsurface finding algorithm and dataset was consisted of several volumes [8]. In[9], lung segmentation was done by extracting lung region by using gray-levelthresholding, and dynamic programming is used to separate both the right and leftregions. Different masks were used to segment the lung region. For lung regionsegmentation, the average of intensity values of Gobar mask, intensity mask, andlung model mask was used [10]. Intensity mask was used to highlight the dark partof the image, and it was the complement of X-ray, and lung model was computedfrom JSRT dataset. To map the model to the input image, bilinear alignment of thelung shape model is used, and a threshold was applied to get a lung region [14].Another technique was proposed in [15] in which morphological operations, cannyedge technique, and Euler number method were used to separate lung region fromthe input image. In [11], a methodology is proposed in which features comprisedof Gabor wavelet transform and fractal dimension are extracted and fuzzy c-meansclustering is used for initial contouring and deformable models based on level setsare used for final contour. Deep learning is also applied on chest radiographs.Convolutional Neural Network(CNN) framework is used for lung segmentation,which consists of seven layers, and it outperformed the manual segmentation [1].Another method is proposed in which Structure Correcting Adversarial Network(SCAN) framework is used. In this method, adversarial process is applied to createsegmentation models for segmentation of chest radiographs [4].

This paper represents a methodology in which kernel mapping is introduced tobring graph cut formulation, which is applied to image segmentation more thanGaussian. Kernel function facilitates image data to map into the higher dimension.Mapping helps piecewise constant model and graph cut formulation to becomeapplicable and also helps to shun the complex modeling of image data. By usinga common kernel function, the minimization of objective functional minimization isperformed by two-step iteration. Minimization is performed with respect to regionparameters and with respect to image partitioning by graph cuts. The proposedtechnique provides the advantage of both simple modeling and optimized graph cut.Before applying segmentation, lung mask is calculated by using training images.Sections II, III, and IV contain methodology, experimentation, and conclusion.

Segmentation of Chest Radiographs for Tuberculosis Screening Using Kernel. . . 27

2 Proposed Methodology

In proposed methodology, lung mask is calculated according to shape similarityof training images with respect to input image, and segmentation is applied forfinding lung boundaries accurately in chest radiographs as shown in Fig. 1. Firstof all, histogram equalization is applied to input as well as remaining images of thedataset. Dataset is trained with their masks and used as training images. Intensityprojections of both input and training images are calculated, and similarity betweentraining images and an input image is measured by using Bhattacharyya coefficientbased on intensity projection. Bhattacharya coefficient is used to detect similaritybetween two samples. Based on Bhattacharyya coefficient, lung mask is calculatedby averaging masks of most five similar images of dataset [2]. To find accuratelung region, segmentation is applied. In proposed segmentation technique, kernelmapping is used to transform image data into the high dimension, and it alsohelps to shun the complex modeling of image data. Segmenting an image is thepartition of image domain into different regions Nseg so that each domain containsdifferent characteristics. Graph cut technique states segmentation of image as anassignment problem. In segmentation, a label is assigned to each pixel of an image.A region is defined as Rl = sε | Ω(s)

.= l. Each region is characterized by onelabel, and graph cut is used to segment the image by finding a label which minimizesthe energy function. The proposed function consists of two terms: kernel-induceddata term and smoothness cost. Data term is defined as MAP formulation usingpiecewise constant model via mapping function. The mathematical representationof segmentation functional is as follows:

F(λ) = D(λ) + βR(λ) (1)

Fig. 1 Flowchart of proposed methodology

28 A. Fatima et al.

D denotes data term and R represents prior. Data term forces the algorithm to assignlabel to pixels with the consistency of image intensities, while smoothness costforces the algorithm to give the same labels to neighboring pixels. Data term canbe defined by using MAP formulation of parametric models as follows:

Fk({ul}, λ) =∑

lεL

∑

sεRl

(Is , ul) + β∑

{s,t}εNr(λ(s), λ(t)) (2)

Usually, different computation models are used to partition images, but these arenot sufficient for complex image models and nonlinear problems. Instead of solvingnonlinear problems, the kernel function is used to transform image data into mappedspace to make piecewise constant model applicable via mapping function and alsohelps to solve linear problems. Mercer’s theorem states that there is no compellingreason to know about φ explicitly [12]. φ(.) supposed to be a mapping of nonlinearfunction from the observation space to high-dimensional space. Each region isassigned by one label. Fk aims to measure the distance of region parameters ul

for 1 ≤ l ≤ Nseg. It can be written as:

Fk({ul}, λ) =∑

lεL

∑

sεRl

(φ(ul) − φ(Is ))2 + β

∑

{s,t}εNr(λ(s), λ(t)) (3)

ul represents the piecewise constant mode parameter of region Rl , and Is containsintensity values of region s. In smoothness term, N represents set which have entirepairs of neighboring pixels and is given by truncated absolute difference. Accordingto Mercer’s theorem, kernel function in high-dimensional space is represented as dotproduct [12]. Instead of knowing mapping explicitly, kernel function can be used asK(x, y), verifying:

K(x, y) = φ(x)T .φ(y), ∀(x, y)εI 2 (4)

where (.) is representing the dot product in higher-dimensional space. Putting kernelfunction in segmentation functional gives:

Jk(Is, μ) = ‖ φ(Is) − φ(μ) ‖2

= (φ(Is ) − φ(μ))T .(φ(Is) − φ(μ))

= φ(Is )T φ(Is) − φ(μ)T φ(Is )

−φ(Is)T φ(μ) + φ(μ)T φ(μ)

= K(Is, Is) + K(μ,μ)

−2K(Ts, μ), με{μl}1≤l≤Nseg (5)


In Eq. (3), kernel function gives the non-Euclidean distance in an image spacecorresponding to squared norm in feature space. Simplification in kernel functiongives the following modified Fk :

Fk({ul}, λ) =∑

lεL

∑

sεRl

Jk(Is , ul) + β∑

{s,t}εNr(λ(s), λ(t)) (6)

The functional depends upon labeling and region parameters in Eq. (6). The mainobjective in segmentation is to minimize functional, and it is done by iterating two-step optimization strategy.

• Update of region parameters• Update of labels

The first step contains fixing label and enhancing Fk with respect to regionparameters {ul}l=1,...,Nreg , and the second step is to find optimal label using graphcut iterations. The algorithm iterates these two steps until convergence.

2.1 Update of Region Parameters

To minimize energy function, region parameters are updated with various iterations.Based on the image domain partition, derivation of Fk is done with respect to regionparameters ul . It gives the following equation:

∂(FK)

∂uk

=∑

sεRk

∂

∂μk

[K(μk,μk) − 2K(Is, μk)] (7)

+β∑

sεCk

∑

tεNs

(μk − μλ(t)) εL (8)

where K represents radial basis function (RBF) kernel and Ck represents pixelslaying on the boundary of region Rk , whereas Ns represents set of neighbors t ofpixel s validating that λ(s) �= λ(t). As radial basis kernel function(RBF) is used inclustering data, the requirement for the minimum of functional Fk with respect touk via fixed point computation gives the following equation.

uk − fRk (uk) = 0, kεL (9)

fRk (uk) =

∑sεRk

(Is, uk) + β∑

sεCk

∑

tεNs

uλ(t)

∑sεRk

(Is, uk) + β∑

sεCk

�Ns

, kεL (10)

30 A. Fatima et al.

The difference of kernel model can be observed evidently in the update of regionparameter in Eq. (9). There would be simple mean update in every region in case ofGaussian model [16].

2.2 Update of Labels via Graph Cut Iterations

The second step consists of graph cuts iteration to get the minimized function oflabeling. Graph cut segmentation can be defined as:

G = 〈v, ε〉 (11)

In Eq. (11), let G be a weighted graph, and V denotes the set of nodes (vertices), andset of edges are denoted by E. Every pixel in an image contains node, and two othernodes source and sink nodes are present in set V . E contains set of edges betweentwo discrete nodes s and t. Cut separate terminals into two induced subgraphs. Themain objective is to find cut with the lowest cost. In the search of minimum functionFk , graph cut iterations are performed on a given subgraph with the given regionparameter provided by the first step. Change in both region parameters and labelscauses a change in weights of the graph [17].

3 Experimentation

3.1 Dataset

Performance of proposed methodology is evaluated on publically available datasetnamed as Japanese Society of Radiological Technology (JSRT). The dataset consistsof 247 CXR images, collected from 14 centers. Each image has 2048 ∗ 2048 pixels.Manual masks of this dataset are also present in which manual left and right lungmasks are generated separately, named as SCR (segmentation in chest radiographs)[20]. Each pixel is represented with 12 bits.

3.2 Results and Discussion

The performance of proposed methodology is evaluated on Dice coefficient. It isused to calculate the overlap measure between the segmented mask and manualmask [6]. Manual masks are generally generated by radiologists. The quantitativeand qualitative results of proposed method can be observed in Figs. 2 and 4,respectively. Figure 2 contains some images segmented from proposed method,


Fig. 2 Visualization of proposed methodology

Fig. 3 Comparison of the segmented mask and ground truth; Column I shows the input image withcontour. Red contour shows segmented result, while yellow contour shows ground mask. ColumnII shows segmented mask, and column III shows ground truth

and Fig. 4 represents values of Dice coefficient among images of JSRT dataset.The alteration in segmented mask and ground truth can be observed in Fig. 3. Incolumn I, red contour shows segmented results, while yellow contour shows manualmasks. Column II is showing the ground truth, and column III is showing thesegmented masks of chest radiographs. Figure 4 represents quantitative performanceon JSRT dataset. X-axis represents the number of images of dataset, and y-axisrepresents Dice coefficient among images of JSRT dataset. Blue dots represent thecoefficient value among images, and the red line shows the average accuracy ofproposed methodology. It is observed that all cases obtained score more than 0.88

32 A. Fatima et al.

250200150100

JSRT Dataset

5000.88

0.89

0.9

0.91

0.92

0.93

Dic

e co

-effi

cien

tDice co-efficient

Performance of Proposed Method on JSRT

Mean

0.94

0.95

0.96

0.97

Fig. 4 Performance evaluation of proposed methodology; x-axis denotes the number of images,and y-axis represents Dice coefficient across JSRT dataset

which is sufficient to detect lung region efficiently. Average accuracy of proposedmethodology of 92.19% ± 0.0377 is achieved.

Dice Co − eff icient (Ω) = | Segmented ∩ Manual || Segmented | + | Manual | (12)

Instead of applying simple graph cut segmentation [2], various iterations of graphcuts are applied, and region parameters are updated in each iteration by using kernelfunction, and it shows more effective as compared to graph cut segmentation on thesame dataset. Table 1 shows the comparison between graph cut and kernel graph cutsegmentation technique.

4 Conclusion

It is difficult to segment chest radiograph due to a large variation in image quality.In this paper, lung mask is obtained by calculating the similarity between inputand training chest radiographs, and then the graph cut segmentation is applied withkernelized energy. In the proposed technique, fixed point iterations are performed


Table 1 Performancecomparison of proposedmethodology and graph cut

Author Accuracy

Candemir et al. [2] 91%

Proposed Method 92.19%

for region parameters, and similarly graph cut iterations are performed for updatinglabels. In this way, the accuracy of 92.19% is achieved as a result.

References

1. Arbabshirani, M. R., Dallal, A. H., Agarwal, C., Patel, A., & Moore, G. (2017). Accuratesegmentation of lung fields on chest radiographs using deep convolutional networks. In SPIEMedical Imaging (pp. 1013305–1013305). International Society for Optics and Photonics.

2. Candemir, S., Jaeger, S., Palaniappan, K., Antani, S., & Thoma, G. (2012). Graph-cut basedautomatic lung boundary detection in chest radiographs. In IEEE Healthcare TechnologyConference: Translational Engineering in Health & Medicine (pp. 31–34).

3. Cootes, T. F., Hill, A., Taylor, C. J., & Haslam, J. (1994). Use of active shape models forlocating structures in medical images. Image and Vision Computing, 12(6), 355–365.

4. Dai, W., Doyle, J., Liang, X., Zhang, H., Dong, N., Li, Y., & Xing, E. P. (2017). SCAN:Structure correcting adversarial network for chest X-rays organ segmentation. arXiv preprintarXiv:1703.08770.

5. Dawoud, A. (2010). Fusing shape information in lung segmentation in chest radiographs.Image Analysis and Recognition, 6112, 70–78.

6. Dice, L. R. (1945). Measures of the amount of ecologic association between species. Ecology,26(3), 297–302.

7. Duncan, J. S., & Ayache, N. (2000). Medical image analysis: Progress over two decades andthe challenges ahead. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(1),85–106.

8. Gill, G., & Beichel, R. R. (2015). Lung segmentation in 4D CT volumes based on robust activeshape model matching. Journal of Biomedical Imaging, 2015, 10.

9. Hu, S., Hoffman, E. A., & Reinhardt, J. M. (2001). Automatic lung segmentation for accuratequantitation of volumetric X-ray CT images. IEEE Transactions on Medical Imaging, 20(6),490–498.

10. Jaeger, S., Karargyris, A., Antani, S., & Thoma, G. (2012 August). Detecting tuberculosis inradiographs using combined lung masks. In 2012 Annual International Conference of the IEEEEngineering in Medicine and Biology Society (EMBC) (pp. 4978–4981). IEEE.

11. Lee, W. L., Chang, K., & Hsieh, K. S. (2016). Unsupervised segmentation of lung fields inchest radiographs using multiresolution fractal feature vector and deformable models. Medical& Biological Engineering & Computing, 54(9), 1409–1422.

12. Muller, K. R., Mika, S., Ratsch, G., Tsuda, K., & Scholkopf, B. (2001). An introduction tokernel-based learning algorithms. IEEE Transactions on Neural Networks, 12(2), 181–201.

13. Naing, W. Y. N., & Htike, Z. Z. (2014). Advances in automatic tuberculosis detection in chestx-ray images. Signal & Image Processing, 5(6), 41.

14. Priyanka, S. M., & Minu, R. I. (2014 February). Improving the conspicuity of lung nodules byuse of” Virtual Dual-Energy” radiography. In 2014 International Conference on InformationCommunication and Embedded Systems (ICICES) (pp. 1–6). IEEE.

15. Saad, M. N., Muda, Z., Ashaari, N. S., & Hamid, H. A. (2014 November). Image segmentationfor lung region in chest X-ray images using edge detection and morphology. In 2014 IEEEInternational Conference on Control System, Computing and Engineering (ICCSCE) (pp. 46–51). IEEE.

34 A. Fatima et al.

16. Salah, M. B., Mitiche, A., & Ayed, I. B. (2010 September). Image partitioning with kernelmapping and graph cuts. In 2010 17th IEEE International Conference on Image Processing(ICIP) (pp. 245–248). IEEE.

17. Salah, M. B., Mitiche, A., & Ayed, I. B. (2011). Multiregion image segmentation by parametrickernel graph cuts. IEEE Transactions on Image Processing, 20(2), 545–557.

18. Van Ginneken, B., & Ter Haar Romeny, B. M. (2000). Automatic segmentation of lung fieldsin chest radiographs. Medical Physics, 27(10), 2445–2455.

19. Van Ginneken, B., Katsuragawa, S., Ter Haar Romeny, B. M., Doi, K., & Viergever, M. A.(2002). Automatic detection of abnormalities in chest radiographs using local texture analysis.IEEE Transactions on Medical Imaging, 21(2), 139–149.

20. Van Ginneken, B., Stegmann, M. B., & Loog, M. (2006). Segmentation of anatomicalstructures in chest radiographs using supervised methods: A comparative study on a publicdatabase. Medical Image Analysis, 10(1), 19–40.

21. Wang, C., Guo, S., & Wu, X. (2009 June). Segmentation of lung region for chest X-ray imagesbased on medical registration and ASM. In 3rd International Conference on Bioinformaticsand Biomedical Engineering, ICBBE 2009 (pp. 1–4). IEEE.

Survey Analysis of Automatic Detectionand Grading of Cataract Using DifferentImaging Modalities

Isma Shaheen and Anam Tariq

1 Introduction

The eye is an important organ of the human body having a complicated system. Itconsists of a number of interconnected subsystems including the lens, pupil, iris,retina, cornea, and optic nerve. There are a number of ocular diseases related todifferent parts of the eye such as glaucoma, trachoma, and age-related maculardegeneration, pathological myopia, retinitis pigmentosa, and diabetic retinopathy[1]. Patients are usually not aware of the gradual progression of their disease.Late diagnosis of ocular diseases makes it difficult to impair vision effectively [1].An ophthalmologist or optometrists diagnose ocular disease after eye examinationthrough slit-lamp or retinal examination or sometimes by visual acuity tests [3].Visual acuity tests are conducted using a chart or a viewing device by reading aseries of progressively smaller letters [3]. A slit lamp allows the ophthalmologistto examine the eye under magnification using intense line of light [4]. In retinalexamination, the pupil is dilated with drops to widen the eye lens to check eye move-ments and papillary responses. These screening methods require expensive medicalequipments that can be used by experienced ophthalmologist. Manual methodsare also time-consuming and subjective based on ophthalmologist experience. Soa number of attempts have been made by the researcher toward the automaticdetection of eye disorders in the past few years. Cataract has caused 47.8% of worldblindness [6]. The greatest risk factor of developing cataract is age. Across the globe,approximately 314 million people are blind or visually disabled due to cataract [6].Among these blinds, 60.7% are needlessly blind because of cataract. More than 90%of these blinds live in developing countries [6]. Cataract is the clouding of the lens

I. Shaheen · A. Tariq (�)College of Electrical & Mechanical Engineering (CEME), National University of Science andTechnology, Islamabad, Pakistan


35


https://doi.org/10.1007/978-3-319-96139-2_4

36 I. Shaheen and A. Tariq

Fig. 1 Normal vs cataract vision [7]

Fig. 2 Formation of cataract as nuclear, cortical, and posterior subcapsular cataract

that occurs when the protein inside the lens clumps together with increasing age.The protein builds up when the older cells compacted into the center of the lenscausing a blur image of the retina. Figure 1 represents the effects of cataract onvision.

Depending on the location and formation of cataract, it has three primary typesincluding nuclear cataract, cortical cataract, and posterior subcapsular cataract [2].Nuclear cataract (NC) is the mostly occurred cataract type due to aging. It isprimarily caused by the hardening and yellowing of the central part of the eye lenscalled nucleus [8]. Cortical cataract (CC) developed in the lens cortex in the formof white wedged-shaped and radially oriented opacities that work from the outsideedge of the lens toward the center in a spoke-like fashion [8]. Posterior subcapsularcataract (PSC) appears as small breadcrumbs or sand particles sprinkled beneath thelens capsule. It usually occurs in steroid and diabetic patients [8]. Figure 2 showsthe formation of these three types of cataract and its effects on eye lens.

Diagnosis of cataract is done by the ophthalmologist after observing the anatom-ical changes occurring in the eye lens. After detection of cataract, grading is done

Survey Analysis of Automatic Detection and Grading of Cataract Using. . . 37

by comparison of slit-lamp image against a set of standard photographs accordingto a grading protocol. Most widely used grading protocols are Lens OpacitiesClassification System III (LOCS-III) [9] and Wisconsin Grading System (WGS)[10].

In this survey we present an overview of methods and techniques developed forcataract detection and grading over the past few years. The publications for thispurpose are taken from PubMed, Springer, and IEEE Xplore. We classify the resultsof review based on the imaging modalities used for detection and grading of cataract.The survey is based on five types of imaging modalities used including slit-lamp,retro-illumination, retinal, digital, and ultrasonic Nakagami images.

2 Automatic Detection and Grading of Cataract UsingSlit-Lamp Images

Slit-lamp images are mainly used for detection of nuclear cataract. As nuclearcataract affects the nucleus of the eye lens, the automatic detection and gradingare done by extracting features from nucleus region [9, 10]. Figure 3 represents theslit-lamp images of nuclear cataract.

There have been number of attempts made toward the automated diagnosissystems of nuclear cataract through slit-lamp photographs. H. Li et al. proposed firstautomatic diagnostic system for nuclear cataract in 2009 [11]. They applied activeshape model (ASM) for lens detection and achieved 95% success rate for featureextraction with 0.36 mean errors for automatic grading. R. Srivastava proposedan image gradient-based approach for grading nuclear cataract using a diverse setof 1700 features and achieved 95% accuracy [12]. X. Gao et al. also proposed adeep learning-based approach for automatic feature learning to grade the severityof nuclear cataract from slit-lamp images [13]. The system performance was betterthan the state-of-the-art clinical grading method yielding 68.6% exact agreementratio and mean absolute error of 0.322 [13]. A.B. Jagadale et al. presented anearly detection and categorization of cataract using circular Hough transform [14].X. Liu et al. proposed a diagnosis framework for pediatric cataract through slit-lamp

Fig. 3 Slit-lamp images with different severity levels for nuclear cataract: (a) normal, (b)immature, (c) mature, (d) hypermature


images [15]. Their system achieved 97.07% mean accuracy, 97.28% sensitivity, and96.83% specificity.

Although the contributions made by the above research work toward automaticdetection and grading of nuclear cataract through slit-lamp imaging help reducethe workload of ophthalmologists, the need of a fully automated system to gradenuclear cataract based on the state-of-the-art grading protocols including LOCS-IIIand Wisconsin grading is still needed.

3 Automatic Detection and Grading of Cataract UsingRetro-illumination Images

Retro-illumination is a non-stereoscopic photograph taken using “Neitz CT-Rcataract camera to focus on the anterior/posterior cortex of the lens” [10]. A retroimage is used to examine cortical and posterior subcapsular cataract. Figure 4 showsthe anterior and posterior cortex retro images representing nonserious and seriouscataract conditions. Some efforts have been made to automate the process of gradingusing retro-illumination images.

H. Li et al. proposed an automatic opacity detection approach for the diagnosisof cortical and posterior subcapsular cataract using retro-illumination images [16].Texture and intensity analysis performed on retro-illumination images improvedROI detection, lens mask generation, and opacity detection [16]. X. Goa et al.

Fig. 4 Retro-illumination images of cortical cataract and posterior subcapsular cataract withdifferent severity levels: (a) normal, (b) immature CC, (c) mature CC, (d) hypermature CC, (e)immature PSC, (f) mature PSC, (g) hypermature PSC


performed pterygium detection on cornea images to enhance the automatic detec-tion and grading of cortical cataract [17]. Their method achieves 85.38% accu-racy with 66.67% specificity and 80.33% specificity for pterygium detection.W. Zhang et al. proposed lens opacity detection approach for serious poste-rior subcapsular cataract using Markov random fields (MRF) and mean gradi-ent comparison (MGC) achieving 91.2% and 90.1% sensitivity and specificity,respectively [18].

Literature survey revealed a very little work done toward the automatic detectionand grading of cortical and posterior subcapsular cataract using retro-illuminationimages. The possibilities of using retro-illumination images for a fully automatedprocess of grading using LOCS-III and Wisconsin grading system still exist.

4 Automatic Detection and Grading of Cataract UsingRetinal Images

Retinal image analysis is widely used for detection of many ocular diseases includ-ing diabetic retinopathy, glaucoma, and cardiovascular and age-related maculardegeneration [19–21]. Retinal images have been rarely used for detection andgrading of cataract, but some researchers made attempts to detect cataract from eyeretinal (fundus) images. Figure 5 represents the retinal images of cataract eye withdifferent severity levels.

M. Yang et al. applied top-bottom hat transformation to improve the classificationof retinal images as cataract or non-cataract [22]. J. Li et al. exploited the usage ofensemble classifiers for automatic detection and grading of cataract through retinalimages. Majority of voting- and stacking-based ensemble methods were used toclassify image as cataract or non-cataract. Empirical experiments showed accuracyup to 93.2% correct classification rate [23, 24]. L. Xiong proposed an approach forcataract diagnosis through retinal images by removing vitreous opacity [25]. Theyevaluated the blurriness of the image and classify it as a five-grade classificationproblem by training a decision tree. The proposed method achieved an accuracy of92.8% with a kappa value of 0.74 between automatic and manual grading [25].

Fig. 5 Retinal images with different severity levels of cataract: (a) normal, (b) immature, (c)mature, (d) hypermature


A. Jamal et al. analyze the retina through vessel detection to categorize oculardiseases [26].

The analysis of automatic detection and grading of cataract using retinal imagesreveal the endless possibilities of using fundus image analysis for detecting cataract.Retinal features such as retinal lesion, blood vessels, optic disc, fovea, and maculacan be helpful for more accurate classification of cataract and non-cataract images.Diagnosis of cataract using retinal images may also improve the diagnosis of otherretinal diseases.

5 Automatic Detection and Grading of Cataract UsingDigital Images

The use of digital camera images for cataract screening is appealing to use takingthe health facilities under consideration in developing countries. Moreover, digitalcamera is simple and easy-to-use equipment as compared to slit lamp and otherextensive medical equipments used for cataract detection. In the last few years, someefforts have been made by the researchers to use digital images in cataract screening(Fig. 6).

R. Supriyanti et al. proposed a simple and robust method for cataract screeningusing front- and backside specular reflection. In case of nonserious cataract condi-tions, both reflections appeared, but for serious cataract conditions, only front-sidereflection appeared, and there is no other reflection that exists due to whitish lenscolor [27]. U. Patwari et al. performed detection, categorization, and assessmentof eye cataract using digital image processing and achieved 94.96% accuracy [28].Y.N. Fuadah et al. used statistical texture analysis and k-nearest neighbor to achievea high accuracy of the system for cataract detection using digital eye images [29].The system achieved an accuracy of 97.2% [29].

Digital/optical images can be used for developing smartphone applications asthese days smartphones are equipped with efficient digital cameras. This wouldhelp the patients in diagnosing their cataract quite earlier even without visiting anyophthalmologist.

Fig. 6 Digital/optical images with different severity levels of cataract: (a) normal, (b) immature,(c) mature, (d) hypermature


6 Automatic Detection and Grading of Cataract UsingUltrasonic Nakagami Images

Ultrasound is a widely used diagnosis method in ocular diseases. Ultrasound A-scan signals are acquired using ultrasound scanner with 30–60 MHz ultrasonictransducers from porcine lenses for cataract diagnosis. Acoustical parametersvelocity, attenuation, and backscattering signals are used to construct B-scan andNakagami images [30]. Figure 7 represents the B-scan and Nakagami images ofdifferent cataract stages from (a) to (c).

Various studies in the past recent years explored the usage of ultrasoundNakagami imaging to quantitatively measure the cataract. P. Tsui et al. explored thefeasibility of using high-frequency ultrasonic Nakagami imaging for characterizingthe cataract lens in vitro [30]. M. Caxinha et al. used ultrasound backscatteringsignals to assess regional cataract hardness [31]. Later on, they presented a newapproach for objective cataract classification using multi-class SVM classifiers[31]. They obtained 97 features in total including acoustical and spectral featuresextracted from the eye lens. Principal component analysis (PCA) was used forfeature selection. Multi-class support vector machines (SVM) are applied forclassification of cataract lens as incipient or advanced cataract and achieved anaccuracy of up to 90.62% [31–33].

Ultrasonic B-scan and Nakagami imaging techniques showed potential forautomatic cataract classification and grading. But the need for a medically provensystem is still required. The developed techniques were tested using porcine eyelens, but the validation from clinical point of view is still need to be done. Acentralized clinical dataset is needed to improve the accuracy of the developedsystems.

Fig. 7 Ultrasonic B-scan images (first row) and Nakagami images (second row) for differentseverity levels of cataract: (a) no cataract, (b) initial cataract, (c) advanced cataract [30]


7 Discussion

Cataract is a diverse disease due to its nature, effects on the eye, varying types,multiple imaging modalities, and diagnostic procedures. The research that has beendone so far limits to one area or the other. Most of the work done is related todetection of nuclear cataract through slit-lamp images and grading using LOCS-IIIor Wisconsin grading protocol. A complete solution for detection and grading ofcortical and posterior subcapsular cataract through a standard grading protocol likeLOCS-III or Wisconsin is still needed to be developed. Literature survey presentedin this paper reveals that there is very little work done for detection of cataract usingretinal image or fundus image analysis. Digital images are a cheap and easy-to-usesolution for automatic diagnosis of cataract. They can be used for early cataractdetection in the rural parts of the country where there is lack of medical facilities.Digital and slit-lamp images can be used in combination to provide a comprehensivesolution for early automatic detection and grading of cataract.

After surgery cataract development is also an important area of study forresearchers and developers working in the field of automatic diagnosis of cataract.As cataract has the chances of developing after the surgery [5, 34], it shouldbe automatically detected whether the lens is implanted or not. It can help theophthalmologist in the surgery of lens.

8 Summary

In Table 1, we further summarize the literature survey analysis of cataract detectionand grading based on different imaging modalities used.

9 Conclusion

Automatic detection and grading of cataract alleviate the burden of ophthalmol-ogists and clinicians. It also provides an objective way to measure the severityof cataract and helps reduce the vision loss by timely and accurate diagnosis. Inthis survey-based paper, we presented an overview of the methods and techniquesdeveloped for cataract detection and gradation. We mainly investigated the usageof four types of imaging modalities used for automatic diagnosis of cataract usingdigital image processing. These types include slit-lamp images, retro-illuminationimages, retinal images, and digital eye images. We also discuss the shortcomings ofthese methods and future research possibilities to improve these methods.


Tab

le1

Sum

mar

yof

lite

ratu

rere

view

S.no

Imag

ing

type

Sam

ple

imag

esE

xist

ing

met

hods

Shor

tcom

ings

1.Sl

it-l

amp

imag

esFe

atur

esex

trac

ted

from

anat

omic

alst

ruct

ure

ofvi

sual

axis

Act

ive

shap

em

odel

(ASM

)B

ag-o

f-Fe

atur

e(B

oF)

mod

elG

roup

spar

sity

regr

essi

on

Full

yau

tom

ated

LO

CS-

III

and

Wis

cons

ingr

adin

gsy

stem

stil

lnee

ded

Inte

grat

ed,f

ully

auto

mat

edso

luti

onfo

rm

ixed

cata

ract

2.R

etro

-il

lum

inat

ion

imag

es

Pupi

ldet

ecti

onus

ing

conv

exhu

llan

dle

ast-

squa

reel

lips

efit

ting

met

hod

Rad

iale

dge

dete

ctio

nal

gori

thm

for

spok

e-li

kefe

atur

eex

trac

tion

Text

ure

anal

ysis

and

imag

ecl

ippi

ngop

acit

yde

tect

ion

[16]

Pter

ygiu

mde

tect

ion

[17]

Full

yau

tom

ated

LO

CS-

III

and

Wis

cons

ingr

adin

gsy

stem

for

grad

ing

ofnu

clea

rca

tara

ctus

ing

clin

ical

feat

ures

Det

ecti

onof

othe

roc

ular

dise

ase

usin

gre

tro-

illu

min

atio

nim

ages

can

beex

plor

ed

3.R

etin

alim

ages

Four

ier

anal

ysis

[22]

Top-

bott

omha

ttra

nsfo

rmat

ion

toen

hanc

eim

age

cont

rast

[22]

Tri

late

ralfi

lter

used

for

nois

ere

duct

ion

Lum

inan

ce-

and

text

ure-

base

dfe

atur

es[2

2]

Ret

inal

feat

ures

like

lesi

on,b

lood

vess

els,

opti

cdi

sc,

fove

a,an

dm

acul

aca

nbe

used

for

cata

ract

dete

ctio

nan

dgr

adin

gD

egre

eof

cata

ract

shou

ldbe

mea

sure

dqu

anti

tativ

ely

4.D

igit

al/o

ptic

alim

ages

Spec

ular

refle

ctio

nap

pear

ance

insi

deth

epu

pil[

16]

Text

ure

unif

orm

ity

and

aver

age

inte

nsit

yin

side

the

pupi

l[16

]

Onl

ytw

ocl

ass

clas

sific

atio

npr

oble

ms

are

addr

esse

dso

far,

i.e.,

cata

ract

and

no-c

atar

act

Deg

rees

ofca

tara

ctst

illn

eed

tobe

auto

mat

ed

5.U

ltra

soni

cN

akag

ami

imag

es

Ult

raso

und

acou

stic

alpa

ram

eter

sve

loci

ty,

atte

nuat

ion,

and

back

scat

teri

ngsi

gnal

sus

edM

ean

inte

nsit

ybr

ight

ness

ofB

-sca

nan

dm

ean

ofN

akag

amip

aram

eter

mis

used

Nee

dto

beve

rifie

dag

ains

tcli

nica

ldat

aset

take

nfr

omre

alpa

tien

tlen

s


References

1. Zhang, Z., Srivastava, R., Liu, H., Chen, X., Duan, L., Wong, D. W. K., Kwoh, C.K., Wong, T. Y., & Liu, J. (2014). A survey on computer aided diagnosis for ocu-lar diseases. BMC Medical Informatics and Decision Making. Retrieved from http://bmcmedinformdecismak.biomedcentral.com/articles/10.1186/1472-6947-14-80

2. Shirole, T. (2014, November 4). About Cataract. Retrieved from http://www.medindia.net/patients/patientinfo/cataract.htm

3. Kinard, E. T. A Closer Look at Cataract. Retrieved from http://www.athenseyecare.net/conditions/cataracts/?

4. All 3 Article What is cataract? (2010, August 4). Retrieved from http://www.parentyourparents.com/pyp_article/cataracts/?

5. Seddon, J., Fong, D., West, S. K., & Valmadrid, C. T. (1995). Epidemiology of risk factors forage-related cataract. Survey of Ophthalmology, 39(4), 323–334.

6. Pizzarello, L., Abiose, A., Ffytche, T., Duerksen, R., Thulasiraj, R., Taylor, H., & Resnikoff, S.(2004). VISION 2020: The right to sight: A global initiative to eliminate avoidable blindness.Archives of Ophthalmology, 122(4), 615–620.

7. Figure 1 Normal Vs Cataract Vision. Retrieved from http://www.eyecenter.com.ph/what-we-do.html#ripen. ©Copyright 2011. American Eye Center.

8. Delcourt, C., Cristol, J. P., Tessier, F., Léger, C. L., Michel, F., Papoz, L., & POLA StudyGroup. (2000). Risk factors for cortical, nuclear, and posterior sub-capsular cataracts: thePOLA study. American Journal of Epidemiology, 151(5), 497–504.

9. Chylack, L. T., Wolfe, J. K., Singer, D. M., Leske, M. C., Bullimore, M. A., Bailey, I. L., Friend,J., McCarthy, D., & Wu, S. Y. (1993). The lens opacities classification system III. Archives ofOphthalmology, 111(6), 831–836.

10. Panchapakesan, J., Cumming, R. G., & Mitchell, P. (1997). Reproducibility of the Wisconsincataract grading system in the Blue Mountains Eye Study. Ophthalmic Epidemiology, 4(3),119–126.

11. Li, H., Lim, J. H., Liu, J., Wong, D. W. K., Tan, N. M., Lu, S., Zhang, Z., & Wong, T. Y.(2009, September). An automatic diagnosis system of nuclear cataract using slit-lamp images.In Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual internationalconference of the IEEE (pp. 3693–3696). IEEE.

12. Srivastava, R., Gao, X., Yin, F., Wong, D. W., Liu, J., Cheung, C. Y., & Wong, T. Y. (2014).Automatic nuclear cataract grading using image gradients. Journal of Medical Imaging, 1(1),014502–014502.

13. Gao, X., Lin, S., & Wong, T. Y. (2015). Automatic feature learning to grade nuclear cataractsbased on deep learning. Biomedical engineering, IEEE transactions on, 62(11), 2693–2701.

14. Jagadale, A. B., & Jadhav, D. V. (2016, April). Early detection and categorization of cataractusing slit-lamp images by hough circular transform. In Communication and Signal Processing(ICCSP), 2016 international conference on (pp. 0232–0235). IEEE.

15. Liu, X., Jiang, J., Zhang, K., Long, E., Cui, J., Zhu, M., An, Y., Zhang, J., Liu, Z., Lin, Z., &Li, X. (2017). Localization and diagnosis framework for pediatric cataracts based on slit-lampimages using deep features of a convolutional neural network. PloS one, 12(3), e0168606.

16. Chow, Y. C., Gao, X., Li, H., Lim, J. H., Sun, Y., & Wong, T. Y. (2011, August).Automatic detection of cortical and PSC cataracts using texture and intensity analysis on retro-illumination lens images. In Engineering in Medicine and Biology Society, EMBC, 2011 annualinternational conference of the IEEE (pp. 5044–5047). IEEE.

17. Gao, X., Wong, D. W. K., Aryaputera, A. W., Sun, Y., Cheng, C. Y., Cheung, C., & Wong, T. Y.(2012, August). Automatic pterygium detection on cornea images to enhance computer-aidedcortical cataract grading system. In Engineering in Medicine and Biology Society (EMBC),2012 annual international conference of the IEEE (pp. 4434–4437). IEEE.

18. Zhang, W., & Li, H. (2017). Lens opacity detection for serious posterior subcapsular cataract.Medical & Biological Engineering & Computing, 55(5), 769–779.

http://bmcmedinformdecismak.biomedcentral.com/articles/10.1186/1472-6947-14-80

http://www.medindia.net/patients/patientinfo/cataract.htm

http://www.athenseyecare.net/conditions/cataracts/?

http://www.parentyourparents.com/pyp_article/cataracts/?

http://www.eyecenter.com.ph/what-we-do.html#ripen


19. Akram, M. U., Tariq, A., Khan, S. A., & Javed, M. Y. (2014). Automated detection of exudatesand macula for grading of diabetic macular edema. Computer Methods and Programs inBiomedicine, 114(2), 141–152.

20. Akram, M. U., Khalid, S., Tariq, A., Khan, S. A., & Azam, F. (2014). Detection andclassification of retinal lesions for grading of diabetic retinopathy. Computers in Biology andMedicine, 45, 161–171.

21. Akram, M. U., Khalid, S., Tariq, A., & Javed, M. Y. (2013). Detection of neovascularization inretinal images using multivariate m-Mediods based classifier. Computerized Medical Imagingand Graphics, 37(5), 346–357.

22. Yang, M., Yang, J. J., Zhang, Q., Niu, Y., & Li, J. (2013, October). Classification of retinalimage for automatic cataract detection. In e-Health Networking, Applications & Services(Healthcom), 2013 IEEE 15th international conference on (pp. 674–679). IEEE.

23. Guo, L., Yang, J. J., Peng, L., Li, J., & Liang, Q. (2015). A computer-aided health-care systemfor cataract classification and grading based on fundus image analysis. Computers in Industry,69, 72–80.

24. Yang, J. J., Li, J., Shen, R., Zeng, Y., He, J., Bi, J., Li, Y., Zhang, Q., Peng, L., & Wang, Q.(2016). Exploiting ensemble learning for automatic cataract detection and grading. ComputerMethods and Programs in Biomedicine, 124, 45–57.

25. Xiong, L., Li, H., & Xu, L. (2017). An Approach to Evaluate Blurriness in Retinal Images withVitreous Opacity for Cataract Diagnosis. Journal of Healthcare Engineering, 2017, 5645498.

26. Jamal, A., Hazim Alkawaz, M., Rehman, A., & Saba, T. (2017). Retinal imaging analysis basedon vessel detection. Microscopy Research and Technique., 80(7), 799–811.

27. Supriyanti, R., & Ramadhani, Y. (2011, June). The Achievement of Various Shapes ofSpecular Reflections for Cataract Screening System Based on Digital Images. In InternationalConference on Biomedical Engineering and Technology (ICBET). Kualalumpur, Malaysia.

28. Patwari, M. A. U., Arif, M. D., Chowdhury, M. N., Arefin, A., & Imam, M. I. (2011). Detection,categorization, and assessment of eye cataracts using digital image processing. In The firstinternational conference on interdisciplinary research and development, 31 May–1 June.

29. Fuadah, Y. N., Setiawan, A. W., & Mengko, T. L. R. (2015, May). Performing high accuracyof the system for cataract detection using statistical texture analysis and K-Nearest Neighbor.In Intelligent Technology and Its Applications (ISITIA), 2015 international seminar on (pp.85–88). IEEE.

30. Tsui, P. H., Huang, C. C., Chang, C. C., Wang, S. H., & Shung, K. K. (2007). Feasibility studyof using high-frequency ultrasonic Nakagami imaging for characterizing the cataract lens invitro. Physics in Medicine and Biology, 52(21), 6413.

31. Caixinha, M., Jesus, D. A., Velte, E., Santos, M. J., & Santos, J. B. (2014). Using ultrasoundbackscattering signals and Nakagami statistical distribution to assess regional cataract hard-ness. Biomedical Engineering, IEEE Transactions on, 61(12), 2921–2929.

32. Caixinha, M., Velte, E., Santos, M., & Santos, J. B. (2014, September). New approachfor objective cataract classification based on ultrasound techniques using multiclass SVMclassifiers. In Ultrasonics Symposium (IUS), 2014 IEEE International (pp. 2402–2405). IEEE.

33. Caxinha, M., Velte, E., Santos, M., Perdigão, F., Amaro, J., Gomes, M., & Santos, J. (2015).Automatic Cataract Classification based on Ultrasound Technique Using Machine Learning: Acomparative Study. Physics Procedia, 70, 1221–1224.

34. Pathak, S., & Kumar, B. (2016). A Robust Automated Cataract Detection Algorithm UsingDiagnostic Opinion Based Parameter Thresholding for Telemedicine Application. Electronics,5(3), 57.

A Privacy Risk Assessment for theInternet of Things in Healthcare

Mahmoud Elkhodr, Belal Alsinglawi, and Mohammad Alshehri

1 Introduction

The Internet of Things (IoT) provides societies, communities, governments, andindividuals with the opportunity to obtain services over the Internet whereverthey are and whenever they want. The IoT enhances communications on theInternet among not only people but also things [6]. It introduces a new conceptof communication, which extends the existent interactions between humans andcomputer applications to things. The IoT has the potential to provide an intelligentplatform for the collaborations of distributed things via local area wireless and wirednetworks and via a wide area of heterogeneous and interconnected networks suchas the Internet [12]. The availability of information coming from nontraditionalcomputer devices in the IoT will shape societies and transform businesses. In 2010,the IoT market value was forecasted to be worth more than 100 billion dollars in2020 [4]. Similarly, in 2013, Cisco forecasted that the economic value created bythe IoT will exceed 14.4 trillion dollars in 2020 [2]. Cisco revised its forecast in2014 to 19 trillion dollars [3] and had been increasingly adjusting its forecast on ayearly basis.

M. ElkhodrCentral Queensland University, Sydney, NSW, Australiae-mail: [email protected]

B. AlsinglawiWestern Sydney University, Sydney, NSW, Australiae-mail: [email protected]

M. Alshehri (�)University of Technology Sydney, Sydney, NSW, Australiae-mail: [email protected]


47





https://doi.org/10.1007/978-3-319-96139-2_5

48 M. Elkhodr et al.

Beyond the massive technological opportunities and benefits the IoT offers,important challenges such as trust, security, and privacy should be considered [8].In the IoT, things, such as sensor devices, will be integrated into streets, homes,work and recreation places, buildings, shopping centres, cars, and other publicenvironments. They will also be carried by people or mounted on mobile vehicles.As a result, things may communicate with each other locally within personal areanetwork (PAN) setups or in a peer-to-peer fashion. They may also interact with IoTapplications remotely over the Internet. In a typical IoT application, IoT devicesmay have the capabilities of automatically sensing, communicating, and processingthe information collected from their environments and their users [14], with ahigh degree of spatial and temporal precision. This information may comprise theexchange of users’ personal and contextual information, including their sensitiveor personal information. Therefore, it is likely that new privacy issues will arisewith such a deep penetration of technology in our life [13]. This paper attemptsto highlight the privacy issues derived from the adoption of the Internet of Thingstechnologies in healthcare. Section 2 discusses the various IoT developments inhealthcare, such as remote health monitoring systems and assistive technologies.The associated and derived privacy issues and challenges are then discussed inSect. 3. This section ends with a brief privacy risk assessment. Concluding remarksare provided in Sect. 4.

2 IoT Development in Healthcare

The healthcare services and communication technology industry have the potentialfor growth in specialised e-health services such as electronic health (e-health),remote monitoring systems, and home and community care among many others[5]. The IoT offers numerous opportunities to improve the operations and delivery ofhealthcare services. The IoT promotes a wider approach to healthcare by addressingthe health needs of a population instead of individuals and by stimulating practicesthat reduce the effects of diseases, disability, and accidental injuries. Additionally,combining healthcare applications with other areas of the IoT stimulates sustainabil-ity in healthcare [15]. It is established in the healthcare community that prevention ofdiseases is as equally important as providing medical treatments [7]. Consequently,the IoT creates the opportunity of maintaining sustainable environments for ahealthier lifestyle.

Other contribution the IoT provides is in reducing the implications of climatechange on the health and well-being of the population [16]. It is essential for thefuture sustainability of healthcare to enable healthcare providers and services tointegrate sustainability principles within their organisations such as energy andwater efficiency and environmental compliance among many others. Also, it iscritical to foster practices that protect and promote the health of communities.Hence, the IoT plays a significant role towards the realisation of a sustainableenvironment, which in turn contributes to a better approach to healthcare. In terms of

A Privacy Risk Assessment for the Internet of Things in Healthcare 49

IoT applications in healthcare, administering medications and the delivery of drugsare among the various envisioned applications in this domain [10].

The integration of IoT technologies in healthcare is expected to result inpromoting remote health monitoring systems as well. Remote health monitoringtechnology provides solutions for monitoring patients at home. These systems aimto deliver higher quality of care and reduce the cost on patients and governmentswithout affecting the quality of the healthcare services provided [5]. The use of aremote monitoring system allows biomedical and other vital signals of a patient tobe measured ubiquitously during his or her daily activities. Such a system allowsthe collection of medical data related to patients’ bodies, such as their heart rates,remotely via the Internet. There are also benefits associated with improving thequality of care and services, such as accuracy and freshness of data obtained andease of accessibility to the patient’s electronic health records (EHRs).

An IoT-based remote monitoring system can detect any changes in the persons’body conditions by monitoring their vital medical signs. The availability andaccessibility of the collected data by this system via the Internet and the abilityto access this EHR, in real time, by various other systems and entities such ashealthcare providers and medical centres, open the door to numerous opportunities[1]. For instance, an alert system can be designed based on analysing the EHRscollected by the remote monitoring system. In the case of a medical emergency, thesystem can be configured to alert the healthcare professionals, emergency services,relatives, and other concerned parties. Also, the system can provide insight into thehealth conditions of a monitored person, so the necessary help can be provided asearly as possible, thus saving lives [1].

On the other hand, applications of IoT in healthcare can be designed to help in themonitoring, early detection, prevention, and treatment of several illnesses [9]. Thisincludes diabetes, heart disease, cancer, seizures, and pulmonary diseases, amongothers. Such diseases usually require constant monitoring of body actions. So theperson needs to be under a constant watch. Traditionally, the medical practitionersand healthcare professionals are responsible for the constant monitoring of patients.However, patients’ monitoring is costly and not as effective as it ought to be.For instance, the doctor is not capable of constantly watching over one patientwith undivided attention. An example of how the IoT can improve patients’monitoring is the integration of Body Sensor Networks (BSN) with other IoT healthsystems [11].

The use of remote monitoring systems could also help in reducing medical errorssince electronic health records (EHRs) are digitally available via the IoT [17]. Theavailability of EHRs makes its retrieval and access more accurate and organised.This will not only help in reducing medical errors but also provide speedy accessto data while maintaining access control privileges as well. IoT applications inhealthcare also extend to personal area networks (PANs). In a PAN, individualsare capable of tracking their bodily functions using various wearable technologiessuch as a wearable smart sensor or a smart watch. Therefore, the adoption of IoT inPAN applications will allow individuals to monitor various aspects of their health,eating and exercise habits, and lifestyle. Examples of these health aspects are the


monitoring of blood pressure, sugar and insulin levels, medicine intakes, heartbeats,sleeping patterns, calorie intake, and others. The capabilities offered by the IoTin this regard are vast. Healthcare professional will be able to access remotelythis information and provide treatment if necessary. This enhances the sharing ofinformation and self-administration of health problems in addition to the earlydetection of diseases.

Henceforth, the ability of accessing health information instantly and remotely viathe IoT enables healthcare professionals to access a new category of information,which was unknown to them before. An example of this information includes thefactors which may have affected the patient’s health, such as their daily routineactivities. Gaining insight into the life of a patient helps in providing a better tailoredtreatment solution. Generally, traditional remote monitoring applications lack theinter-operation that the IoT can provide. Ambient assisted living (AAL) will bepossible with the introduction of an IoT-based system that works concurrently withother IoT applications such as those implemented in smart homes. Additionally,patients will be able to obtain pharmaceutical information regarding the type ofmedicine required instantly and in real time. This includes information about theright dosage, allergy advice, and side effects among others. Not only that, Big Dataanalytic will help in the early detection of diseases before they even develop as well.

3 Privacy: Issues and Open Challenges

In this section, we identify the privacy challenges, risks, and threats facing thedevelopment of IoT in healthcare. As shown in Fig. 1, integrating healthcare servicesin the IoT brings about various cyberattacks to security and privacy. This includestypical network attacks such as replay and snooping attacks among other attacks asillustrated in Fig. 1. Moreover, the proliferation of IoT technologies in the healthcaredomain opens the various open research challenges; these can be summarised to asfollows (Table 1):

– The communications envisioned in the IoT are mainly in the form of machine-to-machine communications. The IoT healthcare devices will progressively performautonomous and complex functions that involve the sharing of personal andsensitive information.

– Many IoT devices are portable or wearable. These devices are characterisedby their low-power and low computational capabilities. Therefore, traditionalcomplex privacy-preserving solutions cannot be implemented in many IoTscenarios.

– The autonomy of IoT devices and how open they become are the main driversfor IoT privacy risks.

– Weak security measures such as insecure networks augment the privacy risks.


ConnectedM-health

Smart medicationEHR

API/Web App

Social EngineeringAttacks

Replay Attacks

Cyber Attacks

WSN attacks

Poison attacks

Exploit / infect

Traffic Sniffing/analysis

TrackingLeakage

Database Cloud Data in Transit Data at rest loT Gateways

Location basedhealth services

Nursing HomesUsers

SmartAmbulance

Connected Hospitals

Smart Homes

IoT healthcare services

Attacker

healthcareProfessionals

Compromised node

Malware

Fig. 1 Vulnerabilities in IoT healthcare

– As with any new technology, there will be an increase in human errors. Therefore,this increase in security hazards will contribute to increasing the privacy risks aswell.

– Vulnerable IoT devices could be used to create botnets of zombie devices thatcan be used to impinge on the privacy of users. Similar to those used to constructDDOS attacks, the botnets in this case will be used to spy on the users. Theywill be used to allow an adversary to easily establish correlations between theuser’s data stored or communicated across various IoT applications and devices.The use of traditional hiding techniques such as anonymisation may prove to beinefficient in this case as well.

– Devices leaking private information can be used as tools for price discrimination.For instance, a business can track a user’s buying habits and patterns and uses thisinformation to generate and raise the price of commodities at any given time.

– Complex encryption techniques will not make sense for all IoT devices. There-fore, there is a need for a framework to share private information under a privateumbrella.

– How to implement traditional privacy principles such as data minimisation, need-to-know, and informed consent in an automated, dynamic, and highly scalableIoT environment remains an open challenge.


Table 1 Examples of attacks on IoT healthcare systems

Threat Description of attack and threatagent

Severity

Data theft Infrastructureattacks

Any connected medical device is atrisk to be exploited by malicioususers

Medium

Worm, phishing attacks Vulnerable and weak devices canbe used as a back door to propagateworm attacks

High

Availability attacks on criti-cal infrastructure

Attacks on monitoring, emergencyand monitoring devices and sys-tems

High

Stored private data areleaked

Insecure communications TypicalMITM attacks CompromisedEHRs databases Malicioususers/nodes

Medium

Private data are unintention-ally leaked

Human errors typical Hazardattacks

Low

Poison and integrity attacks For example, exploiting prescrip-tions or drug libraries and databases

High

Insecure device Many IoT sensor devices have low-power and low capabilities (strongencryptions cannot be used)

High

Network attacks (DoS,DDOS, spoofing, etc.)

Traditional network attacks DoS,attacks on routing, Smurf attacks,etc.

High

WSN-inspired attacks IoT networks are also vulnerableto sinkhole, exhaustion, overlap,flooding attacks, and aggregationattacks among others

Med

Inference attack Data mining and analysing datacollected from various medicaldatabases and IoT networks or viareconnaissance attacks

Low

– Due to the increase in connectivity in the IoT, many IoT applications will betempted to collect users’ personal or sensitive information outside their scope,i.e., mission creep.

– Accountability is a major issue in the IoT as well. The IoT may open the doorto a new vector of cyberattacks. How to handle issues relating to the violation ofindividuals’ privacy? Are data protection laws up to date with the technology?And how to enforce these laws are also open challenges.

– When an IoT network or application becomes global or universal across statesand countries, which data protection laws will apply?

– How to ensure the right to be forgotten?


– How to prevent adversaries from exploiting the vulnerabilities of IoT devices andcollect the private information of individuals for cyberbullying purposes such asrevenge porn?

Although privacy has been always a concern in the Internet world, the IoT bringsadditional challenges to the protections of the users’ private and sensitive informa-tion. Henceforth, the development of robust privacy-preserving techniques plays avital role towards the proliferation of the IoT. Various privacy protection methodshave been proposed in the literature to deal with privacy issues. However, mostof these protection techniques are designed to work within traditional computerenvironments. Other techniques designed for WSNs, which typically cater for thelow-cost and low-power requirements of sensors, do not consider the heterogeneity,scalability, and autonomy of communications provisioned by the IoT. Therefore,there is a need for self-adaptable middleware solutions in the IoT that balancebetween the privacy of users and the increase in access demands to the users’generated data.

4 Conclusion

The applications of IoT in healthcare span from those used in hospitals, medicalcentres, nursing homes, and smart homes to those used as part of a personal areanetwork. The IoT has the potential for further advancements and innovations inseveral other areas such as early detection of illnesses and their prevention. The IoTis shaping modern healthcare with promising technological, economic, and socialopportunities. This paper presented advances in IoT in the areas of remote healthmonitoring systems and touched upon other IoT-based healthcare technologies.The paper then concluded with a discussion on the pressing issues challenging theprivacy of users in the IoT paradigm specifically in the healthcare domain.

References

1. Alsinglawi, B., Elkhodr, M., Nguyen, Q. V., Gunawardana, U., Maeder, A., & Simoff,S. (2017). Rfid localisation for internet of things smart homes: A survey. arXiv preprintarXiv:1702.02311.

2. Bradley, J., Barbier, J., & Handler, D. (2013). Embracing the internet of everything to captureyour share of 14.4 trillion. White Paper, Cisco Systems Inc 17

3. Bradley, J., Reberger, C., Dixit, A., & Gupta, V. (2014). Internet of everything: A 4.6 trillionpublic-sector opportunity. White Paper, Cisco Systems Inc 17

4. Kecskemeti, G., Casale, G., Jha, D.N., Lyon, J. and Ranjan, R. (2017). Modelling andsimulation challenges in internet of things. IEEE cloud computing, 4(1), pp. 62–69. https://ieeexplore.ieee.org/abstract/document/7879128/

5. Elkhodr, M., Shahrestani, S., & Cheung, H. (2011). Ubiquitous health monitoring systems:Addressing security concerns. Journal of Computer Science 7(10), 1465–1473

https://ieeexplore.ieee.org/abstract/document/7879128/

https://ieeexplore.ieee.org/abstract/document/7879128/


6. Elkhodr, M., Shahrestani, S., & Cheung, H. (2016). A middleware for the internet of things.arXiv preprint arXiv:1604.04823.

7. Glasgow, R. E., Vogt, T. M., & Boles, S. M. (1999). Evaluating the public health impact ofhealth promotion interventions: The re-aim framework. American Journal of Public Health,89(9), 1322–1327.

8. Gubbi, J., Buyya, R., Marusic, S., & Palaniswami, M. (2013). Internet of things (IoT): Avision, architectural elements, and future directions. Future Generation Computer Systems,29(7), 1645–1660.

9. Hiremath, S., Yang, G., & Mankodiya, K. Wearable internet of things: Concept, architecturalcomponents and promises for person-centered healthcare. In: 2014 EAI 4th InternationalConference on Wireless Mobile Communication and Healthcare (Mobihealth) (pp. 304–307).IEEE.

10. Hu, F., Xie, D., & Shen, S. On the application of the internet of things in the field of medical andhealth care. In: Green Computing and Communications (GreenCom), 2013 IEEE and Internetof Things (iThings/CPSCom), IEEE International Conference on and IEEE Cyber, Physicaland Social Computing (pp. 2053–2058). IEEE.

11. Lo, B. P., Ip, H., & Yang, G. Z. (2016). Transforming health care: Body sensor networks,wearables, and the internet of things. Published in: IEEE Pulse 7(1), 4–8.

12. L. Mainetti, L. Patrono and A. Vilei, (2011). Evolution of wireless sensor networks towardsthe Internet of Things: A survey, SoftCOM 2011, 19th International Conference on Software,Telecommunications and Computer Networks, Split, pp. 1–6. https://ieeexplore.ieee.org/document/6064380/

13. Medaglia, C. M., & Serbanati, A. (2010). An overview of privacy and security issues in theinternet of things (pp. 389–395). New York: Springer.

14. Sicari, S., Rizzardi, A., Grieco, L. A., & Coen-Porisini, A. (2015). Security, privacy and trustin internet of things: The road ahead. Computer Networks, 76, 146–164.

15. Turcu, C. E., & Turcu, C. O. (2013). Internet of things as key enabler for sustainable healthcaredelivery. Procedia-Social and Behavioral Sciences, 73, 251–256.

16. Vermesan, O., Friess, P., Guillemin, P., Gusmeroli, S., Sundmaeker, H., Bassi, A., Jubert, I.S., Mazura, M., Harrison, M., & Eisenhauer, M. (2011). Internet of things strategic researchroadmap. Internet of Things-Global Technological and Societal Trends, 1, 9–52.

17. Zhi-peng, Y. (2012). Application of internet of things in community health service. Internet ofThings Technologies, 9, 026.

https://ieeexplore.ieee.org/document/6064380/

https://ieeexplore.ieee.org/document/6064380/

Parallel Computation on Large-ScaleDNA Sequences

Abdul Majid, Mukhtaj Khan, Mushtaq Khan, Jamil Ahmad, Maozhen Li,and Rehan Zafar Paracha

1 Introduction

Deoxyribonucleic acid (DNA) is an important molecule that holds all geneticinformation and “instructions” for an organism. The human genome is composedof over 3 billion base pairs of information organized into 23 chromosomes [1].DNA analysis is important for discovery of differences and similarities of organismsand exploration of the evolutionary relationship between them. This process oftenrequires comparisons of the corresponding DNA sequences, for example, checkingwhether one sequence is a subsequence of another or comparing the occurrences ofa specific subsequence in the corresponding DNA sequences [2].

DNA sequence comparison and identification techniques facilitate the extractionof significant information from reported data. It can be used to infer the evolutionaryhistory and biological function of the sequence as a query. Comparison among thequery and known DNA/RNA sequences is becoming a major tool for phylogeneticanalysis, drug design, evolution, biodiversity, epidemiology, pharmacogenomics,

A. Majid · M. Khan (�)Department of Computer Science, Abdul Wali Khan University Mardan, Mardan, Pakistane-mail: [email protected]

M. KhanPakistan Ordinance Factory (POF), Wah Cantt, Pakistan

J. Ahmad · R. Z. ParachaDepartment of Computational Sciences, National University of Science and Technology,Islamabad, Pakistane-mail: [email protected]; [email protected]

M. LiDepartment of Electronic and Computer Engineering, Brunel University, Uxbridge, UKe-mail: [email protected]


55






https://doi.org/10.1007/978-3-319-96139-2_6

56 A. Majid et al.

and detection of genetic diseases. A number of techniques have been proposedfor DNA analysis in [3–8]. However, these techniques are sequential and thereforeunable to process large-scale DNA sequences within reasonable time.

With rapid development in bioinformatics and computational biology, the col-lected DNA dataset is growing exponentially, doubling every 18 months [9]. Dueto large-scale and complex structure of the DNA dataset, the analyses of DNAsequences are becoming computationally a challenging issue in bioinformatics fieldand computational biology. A fast, sophisticated, and parallel computing approachis required in the field of bioinformatics and computational biology that providesthe capabilities to analyze large amount of DNA sequences within reasonable time.For this purpose, a number of parallel techniques have been proposed in [10–13] inorder to deal with the large-scale DNA analysis computation requirement.

Vector space model (VSM) [14–17] is a widely used approach in informationretrieval system. VSM considers the documents as vectors in n-dimensional space(where n is the number of distinguishing terms used to describe contents of thedocuments), and vectors are compared for similarity by computing the cosine valueamong them [3, 15]. The documents are retrieved based on the measurement of thesimilarity between the query and the documents. This means that documents witha higher similarity to the query are judged to be more relevant to it and should beretrieved by information retrieval system in a higher position in the list of retrieveddocuments.

VSM technique is also applied in the field of bioinformatics and computationalbiology for the analysis of biological datasets. Sarkar et al. [18] proposed VSM-based approach that identifies potential relationship between complex diseasesbased on DNA sequence similarity. The approach was employed on two differentdiseases such as Alzheimer’s and Prader-Willi syndrome. Similarly Abdul-Rub etal. [6] proposed a modified VSM for the analysis of protein datasets. The authorsmodified the original VSM method in order to work with protein dataset. Themodified VSM showed good results in terms of accuracy; however, it takes highcomputation time for processing massive collections of protein sequences.

In this paper we present a parallel VSM approach which is built on top ofmodified VSM [6] that analyzes large-scale DNA datasets for similarity search.The main difference between the proposed PVSM and modified VSM is that thePVSM is a parallelized computation on a number of available processing cores tosupport processing of massive volume of DNA sequences, whereas the modifiedVSM approach is working only on single processing core and limited to processlarge-scale DNA sequences. The proposed approach determines similarity betweenthe given documents and user query document based on degree of relevance(i.e., cosine value). The performance of the PVSM is evaluated using variedsizes of DNA sequences from aspect of computation efficiency and accuracy. Theevaluation results show that the PVSM significantly outperforms the modified VSMin computation efficiency while maintaining same level of accuracy of the modifiedVSM.

The remainder of the paper is organized as follows: Section 2 introduces vectorspace model. Section 3 presents the design of parallel VSM. Section 4 evaluates the

Parallel Computation on Large-Scale DNA Sequences 57

performance of parallel VSM. Finally, Sect. 5 concludes the paper and points outsome future work.

2 Vector Space Model

Information retrieval system is the process of representation, storage, organization,and access to information items [11]. In information retrieval process, informationare retrieved according to the degree of relevance (similarity) between the searcheddocuments and user query [12, 13]. There are many approaches proposed for theinformation retrieval system, among them one approach is vector space model(VSM) [3, 14].

Vector space model is the most widely used technique for information retrievalapplications due to its simplicity and efficiency over large document collections,and it is very appealing to use. One of the main advantages of VSM is that itsimplementation is simple even to deal with millions of documents. The VSMis constructed on two sets of calculations which are based on term geometrywhereby each term has its own dimension in a multidimensional space, queries,and documents. These calculations are:

First, it calculates the weight of every term in query document and in inputdocuments. This calculation determines the magnitude and importance of the termin the input documents.

Second, it calculates the cosine value by comparing the query vector with thevector of each input document. This calculation shows the similarity between thequery and input documents.

The weight of term can be calculated using Eq. (1):

wi = tf i × log

(D

dfi

)(1)

where tfi is the term frequency, “D” is the number of documents, and dfi is thedocument frequency of term “i.” Log is used to reduce the effect relative to termfrequency.

The cosine value can be computed using Eq. (2):

Sim (Q,Di) =∑i

WQ,jWi,j

√∑j

W 2Q,j

√∑i

W 2i,j

(2)

where “Q” is the query and Di is the ith document. WQ, j is the weight of the term“j” in query Q, and Wi, j is the weight of the term “j” in document “i.”

58 A. Majid et al.

3 Design of Parallel Vector Space Model

In this section we present the proposed parallel VSM for large DNA data analysis.The proposed technique is developed on top of the modified VSM [6]. Themodified VSM is sequential computation approach which takes huge execution timewhen processing large-scale DNA sequences, whereas the PVSM is parallelizedcomputations on a number of processing cores which significantly reduce the overallexecution time. The parallel VSM model performs the following steps for DNAsimilarity search.

3.1 Splitting Query and Documents into Terms

In the first step, PVSM splits the query sequences and input document sequencesinto terms of desired size and stores them into lists or vectors for matching betweenquery and input document sequences. The process of this step is shown in Fig. 1. InFig. 1, D1 and D2 are input documents which are parsed in parallel. LP1 and LP2are logical processor 1 and logical processor 2, respectively.

3.2 Finding Terms Frequency

In the second step PVSM finds the terms frequency (occurrence) of query stringin query and given documents through matching process. One logical processorhandles one matching process which means that multiple logical processors canperform parallel matching in a single iteration. The process of this step is shown inFig. 2.

Fig. 1 Splitting documents into term size


Fig. 2 Finding term frequency

For the above scenario, each LP takes one term (e.g., CTCT) and finds theoccurrence of the term in Query, Doc1, and Doc2 simultaneously in a singleiteration. It means that operation sets {1,2,3,4}, {5,6,7,8}, and {9,10,11,12} areexecuted in parallel to match the respective terms in Query and documents. Insequential manner, the first term “CTCT” will be picked for matching with Queryand Documents (Doc1 and Dco2) in first iteration. Then “GAGG” will be matchedin the second iteration, and finally matching of “GGGT” will be performed in thethird iteration.

3.3 Calculation of Terms Weight

In the third step PVSM calculates the weight of each term using Eq. (1). The weightof a term reflects the importance of the term in given documents and use in thecalculation of vector length, dot product, and cosine value. One logical processor isused to calculate the weight of a term. Hence, PVSM calculates weight of multipleterms in parallel using all available processing cores as shown in Fig. 3.

In Fig. 3, LP-1, LP-2, LP-3, and L P-4 are the logical processors, “D” is the totalnumber of document sequences, and “DFI” is the number of document sequenceswhere a specific term appears.

3.4 Calculation of Cosine Values

Finally, PVSM computes cosine values of each document in parallel as shown inFig. 4. The cosine value determines how much similar a particular document is to

60 A. Majid et al.

Fig. 3 Calculation of termsweight

Fig. 4 Calculates cosinevalue

the query. The higher the cosine value of a document, the more is the similarity ofthe query sequence to that document sequence.

In figure 4, D1, D2, D3, and D4 are represented documents. LP-1, LP-2, LP-3, and LP-4 are logical processors. “DP” represents dot product of both query anddocument. “VL(Q) and VL(Di)” are vector lengths of query “Q” and document“Di,” respectively.

4 Evaluation and Experimental Results

We have compared the performance of the PVSM with that of the sequential VSMfrom the aspects of both efficiency in computation and accuracy. The performancewas evaluated using different number of DNA sequences.

In this section first we give a brief description on the experimental environmentthat we used in the performance comparison process and then discuss the experi-mental results.


Table 1 Dataset details

Official symbol TNFAIP2Official full name TNF (tumor necrosis factor) alpha-induced protein 2Gene type Protein codingLineage Eukaryota; Metazoa; Chordata; Craniata; Vertebrata; Euteleostomi;

Mammalia; Eutheria; Euarchontoglires; Primates; Haplorrhini;Catarrhini; Hominidae; Homo

Number of sequencestested for matching

200

4.1 Experimental Environment

The experiments were performed on Intel Core M-5Y10c CPU having two physicalcores and 4 x logical processors with 2GB memory. The proposed algorithm wasimplemented in C#.NET 4.0. The Task Parallel Library of C#.NET was used inorder to achieve task parallelism. The analysis of the sequential VSM was carriedout on one processor, whereas the PVSM was run on up to four logical processors.In order to address the difference in result measurements, we have performed eachexperiment ten times and took average execution time.

For the analysis of both sequential VSM and PVSM, the DNA datasets wereselected from NCBI [19]. We used two datasets, i.e., a query dataset which isused as a query and an input dataset that contained multiple documents of theDNA sequences. Hence we selected TNFAIP2 (tumor necrosis factor, alpha-inducedprotein 2) DNA sequence of Homo Sapiens as a query sequence, and the sameTNFAIP2 DNA sequences of different species are selected as input dataset witheach sequence containing approximately 5000–34,000 bases. The specific details ofthe datasets used in the experiments are displayed in Table 1.

4.2 Results

A number of experiments were carried out to evaluate the efficiency and accuracyof the PVSM implementation. The sequential VSM was implemented following themethod mentioned in [6] on single processor, whereas the PVSM was parallelizedon four logical processors. From Fig. 5, it can be observed that PVSM significantlyperformed better than sequential VSM in computation using four logical cores.For example, the sequential VSM takes 11.86 s, 29.88 s, 64.06 s, and 249.85 swhen processing 50, 100, 150, and 200 DNA sequences, respectively, belonging todifferent species, whereas the PVSM takes 10.26 s, 22.42 s, 44.13 s, and 125.33 swhen processing the same number of sequences, respectively.

As the PVSM algorithm parallelized the computation on a number of processors,there may be a possibility that the parallelization approach affects the accuracy levelof PVSM. Hence, we analyzed the accuracy of PVSM by comparing the resultant

62 A. Majid et al.

500

50

100

150

200

250VSM Execution Time

PVSM Execution TimeE

xecu

tion

Tim

e(s)

100

Number of DNA Sequences

150 200

Fig. 5 Computation efficiency analysis of PVSM

500.86

0.88

0.9

0.92

0.94

0.96

0.98

1VSM

PVSM

100 150 200Number of DNA Sequences

Cos

ine

Val

ue

Fig. 6 Accuracy analysis of PVSM

cosine values of both PVSM and sequential VSM as shown in Fig. 6. Table 2 clearlydemonstrates the resultant cosine value of both the PVSM and sequential VSM.From Table 2, it can be seen that both PVSM and sequential VSM calculated thesame cosine value for the respective datasets. Hence the parallelization approachdoes not affect the accuracy level of the PVSM.

We evaluate the performance of PVSM in terms of speedup when processingvaried number of DNA sequences. In order to compute the speedup, we use (3):


Table 2 Accuracy analysis of PVSM

Number DNA sequences Cosine of VSM Cosine of PVSM

50 0.98962 0.98962100 0.99113 0.99113150 0.98964 0.98964200 0.87275 0.87275

501.1

1.2

1.3

1.4

1.5

1.6

1.7

1.8

1.9

2Speedup of PVSM

Spe

edup

100 150 200Number of DNA Sequences

Fig. 7 Speedup analysis of the PVSM

Speedup = T dS

T dN

(3)

where T dS is execution time of the sequential VSM on single processor when

processing d number of sequences. The value of dis 50, 100, 150, and 200. T dN is the

execution time of PVSM parallelized on Nnumber of processors when processingd number of sequences. In our case the value of N is 4. The results of Eq.(3) aredisplayed in Fig. 7. The PVSM performed best in speedup on large number of DNAsequences with 200 sequences. It can be observed that the speedup of the PVSM isincreased with increasing number of sequences. For example, when processing 50numbers of DNA sequences, the PVSM achieved 1.16 times speedup, whereas thePVSM generated 1.99 times speedup when processing 200 numbers of sequences.The PVSM achieved maximum speedup on a large number of sequences whichshows that the PVSM method is highly applicable for processing large numbers ofDNA sequences.

64 A. Majid et al.

5 Conclusion and Future Work

In this paper we have presented PVSM, a parallel vector space model that analyzeslarge-scale DNA sequences taking advantage of a multi-core system. The proposedapproach was parallelized on a multi-core system using C# Task Parallel Librarywhich is included in NET 4.0. The experiment results have shown that the PVSMperform better than the sequential VSM in computation efficiency while maintainingthe similar level of accuracy in comparison with the sequential VSM.

The proposed technique was parallelized on a small number of logical cores(i.e., 4). One future work would be to test the performance, i.e., computationefficiency and scalability of the proposed approach, on a large number of pro-cessing cores. The sequence term size can impact on accuracy and computationefficiency of the proposed method. So another future work would be to evaluatethe performance of the proposed technique from the aspect of varied sizes ofsequence terms.

References

1. Bald, P., Baronio, R., Cristofaro, E. D., Gasti, P., & Tsudik, G. (2000). Efficient and securetesting of fully-sequenced human genomes. Biological Sciences Initiative, 470, 7–10.

2. Memeti, S., & Pllana, S. 2016. Analyzing large-scale DNA sequences on multi-core archi-tectures. Proceedings – IEEE 18th international conference on computational science andengineering CSE 2015, pp. 208–215.

3. Ogheneovo, E. E., & Japheth, R. B. (2016). Application of vector space model to query rankingand information retrieval. International Journal of Advanced Research in Computer Scienceand Software Engineering, 6(5), 42–47.

4. Smith, T. F., & Waterman, M. S. (1981). Identification of common molecular subsequences.Journal of Molecular Biology, 147(1), 195–197.

5. Dereeper, A., Audic, S., Claverie, J.-M., & Blanc, G. (2010). BLAST-EXPLORER helps youbuilding datasets for phylogenetic analysis. BMC Evolutionary Biology, 10(1), 8.

6. Abual-Rub, M., Abdullah, R., & Rashid, N. (2007). A modified vector space model for proteinretrieval. International Journal of Computer Science and Network Security, 7(9), 85–89.

7. Patel, S., Panchal, H., & Anjaria, K. (2012). DNA sequence analysis by ORF FINDER amp;GENOMATIX tool: Bioinformatics analysis of some tree species of Leguminosae family, in2012 IEEE international conference on bioinformatics and biomedicine workshops, pp. 922–926.

8. Vandin, F., Upfal, E., & Raphael, B. J. (2012, March). Algorithms and Genome Sequencing :Identifying Driver Pathways in Cancer. IEEE Computer Magazine, 45(3), 39–46.

9. Benson, D. A., Cavanaugh, M., Clark, K., Karsch-mizrachi, I., Lipman, D. J., Ostell, J., &Sayers, E. W. (2013). GenBank. Nucleic Acids Research, 41(D1 November 2012), 36–42.

10. de Almeida, T. J. B. M., & Roma, N. F. V. (2010, February). A Parallel ProgrammingFramework for Multi-core DNA Sequence Alignment, 2010 international conference onComplex, Intelligent and Software Intensive Systems (CISIS), 2010, pp. 907–912.

11. Marçais, G., & Kingsford, C. (2011). A fast, lock-free approach for efficient parallel countingof occurrences of k-mers. Bioinformatics, 27(6), 764–770.

12. Herath, D., Lakmali, C., Ragel, R. (2012, March). Accelerating string matching for bio-computing applications on multi-core CPUs. IEEE 7th, Int. Conf. Ind. Inf. Syst. ICIIS 2012.


13. Takeuchi, T., Yamada, A., Aoki, T., & Nishimura, K. (2016). cljam: A library for handlingDNA sequence alignment/map (SAM) with parallel processing. Source Code for Biology andMedicine, 11, 1–4.

14. Manning, C. D., Raghavan, P., & Schütze, H. (2008), An introduction to information retrieval,Cambridge University Press, 2008.

15. Raghavan, V. V., & Wong, S. K. M. (1986). A critical analysis of vector space modelfor information retrieval. Journal of the American Society for Information Science, 37(5),279–287.

16. Singhal, A. (2001). Modern information retrieval : A brief overview. IEEE Data EngineeringBulletin, 24, 35–43.

17. Castells, P., Fernandez, M., & Vallet, D. (Feb. 2007). An adaptation of the vector-spacemodel for ontology-based information retrieval. IEEE Transactions on Knowledge and DataEngineering, 19(2), 261–272.

18. Sarkar, I. N. (2012). A vector space model approach to identify genetically related diseases.Journal of the American Medical Informartion Association, 19(2), 249–254.

19. “NCBI,” National Center for Biotechnology Information. [Online]. Available: https://www.ncbi.nlm.nih.gov/. Accessed 26 Jan 2017.

https://www.ncbi.nlm.nih.gov

Augmented and Virtual Reality in MobileFitness Applications: A Survey

Ryan Alturki and Valerie Gay

1 Introduction

Obesity is a major health problem around the world. Obesity can be defined asabnormal or excessive fat accumulation that may impair health [1]. Around theworld, 15% of the current population are considered to be obese, and almost 40%of the current population are suffering from overweight [2]. Both obesity andoverweight are seen as the main reason for several dangerous chronic diseases, forexample, diabetes and hypertension [3, 4]. As a result of this, several researcherswere motivated to find a way to control and stop the spread of obesity [5–8].The majority of researchers’ results concluded that obesity could be controlledand stopped by doing physical exercises and changing eating habits. Nevertheless,several experts believe that to motivate people who are suffering from obesity tolose weight and have a better lifestyle is not easy. They believe that behaviourintervention is seen as one of the best ways for changing behaviour that is related tofitness and health [9, 10].

During the last few years, the use of fitness mobile apps is becoming populararound the world and especially by people suffering from obesity and want to loseweight to have a better lifestyle. According to a recent study, the international fitnesstechnology market is estimated to be worth around 19 billion US dollars in 2014[11]. According to a report undertaken by Nielsen’s Mobile NetView, one-third(46 millions) of US smartphone owners use fitness mobile apps [12]. Around theworld, 16% of all Internet users use health and fitness mobile apps [13]. In 2014,

R. Alturki · V. Gay (�)Faculty of Engineering and Information Technology, University of Technology Sydney, Ultimo,Australiae-mail: [email protected]; [email protected]


67




https://doi.org/10.1007/978-3-319-96139-2_7

68 R. Alturki and V. Gay

the mobile app business had expanded by 15% overall. Regarding average dailyusage, the number of health and fitness apps increased by 62% [14].

Supporting health behaviour change via using mobile fitness and health apps ispromising. The number of fitness and health apps has increased rapidly in the lastfew years, and by today, there are more than 31,000 fitness and health apps availableto use [15]. Moreover, the interest of how fitness’ apps role can influence thebehaviour of people who suffer from obesity is growing. A recent study introduceda framework that named “Functional Triad” which aims to describe the device’srole in the device-human interaction [16]. The study explains that devices can actas tools, mediums and social actors for motivating human. For instance, fitnessmobile apps can play the role of predisposing tools for diffusing fitness information.Furthermore, personal information regarding users’ behaviour can be collected bythem and can connect users to several social networks.

To influence the behaviour of obese individuals, fitness mobile apps should haveunique features that play an important role in motivating obese individuals. A recentsurvey states that there are four main features that motivate obese individuals to usefitness mobile apps in order to lose weight and have a better lifestyle. These featuresare [17]:

• Goal settings• Monitoring, tracking and feedback• Reminders and alerts• Rewards or gamification

However, there are several technologies that have emerged over the last year andhad been used widely in mobile apps as a motivational tool, for example, augmentedreality (AR) and virtual reality (VR). Pokémon Go is a mobile app game that usesAR technology as an entertainment tool. According to a recent report, Pokémon Gohas been downloaded 650 million times around the world [18].

This research seeks to contribute to significant researches concerning both AGand VR technologies in mobile apps. The aim of this study is to conduct a systematicreview which reveals the most prominent and recent AG and VR studies in mobileapps that have been discussed and have emerged in the literature. This researchis going to be useful for developing a fitness mobile app that considers both AGand VR technologies. This survey is unique because it discusses some of the mostcontemporary literature.

2 The Systematic Review

We undertook a systematic review to search for published, peer-reviewed articlesthat investigated AG and VR in mobile apps. We utilised the terminology outlinedin the table below (Table 1) to look for research papers covering AG and VR inmobile devices and applications. We sought to incorporate all the related terms thatcould provide us with articles relevant to this topic.

Augmented and Virtual Reality in Mobile Fitness Applications: A Survey 69

Table 1 Keywords used in the systematic review relating to AG and VR technologies in mobileapplications

Search lines Search terms Filtered by

Line 1 Mobile device or mobile phone or smartphone Title/abstract2. AND Applications or apps Title/abstract3. AND Fitness applications or fitness apps Title/abstract4. AND Augmented reality or AR Title/abstract5. AND Virtual reality or VR Title/abstract6. AND Augmented reality in mobile applications or AR in

mobile apps or AR in appsTitle/abstract

7. AND Virtual reality in mobile applications or VR inmobile apps or VR in apps

Title/abstract

We referred to JMIR, CINAHL, Academic Search Premier, PsycINFO, HealthSource, Communication and Mass Media Complete, Computers and AppliedSciences Complete, Psychology and Behavioural Sciences Collection, ComputerSource, PubMed, Web of Science and PsycARTICLES.

We have followed the methodology from [17], and the flow chart below showshow the systematic review was undertaken (Fig. 1).

3 Results

We conducted a systematic literature review of AR and VR in mobile apps. Welooked for articles that discussed AR and VR in mobile devices and applications.We also conducted a comprehensive literature review on AR and VR in mobile appsand tried to figure out the important attributes discussed in these papers.

3.1 Augmented Reality in Mobile Apps

In AR, physical reality can become enhanced through the additional informationthat computers can generate in real time [19]. Over the time, the definition of ARhas been broadened, and the following properties are believed to be part of any ARsystem [20]:

• AR systems combine virtual and real objects in a real environment.• AR systems run in real time and interactively.• AR systems align or register virtual and real objects with each other.

Milgram and Kishino introduced a continuum of real-to-virtual environments.In the continuum, they showed AR as part of mixed reality. The surroundingenvironment in AR is always real unlike augmented virtuality and virtual envi-


Records identified through databasesearching

Records after duplicates removed

n=97

n=62

n=51

n=36

n=24

Records after initial screening of titleand abstract

Records excluded onthe basis of title andabstract n=15

Final exclusion (Non-mobile studies,qualitative studies)n=12

Full text articles accessed for finalinclusion

Studies included for qualitativesynthesis

Incl

uded

Elig

ibili

tyS

cree

ning

Iden

tific

atio

n

Fig. 1 Methodology for the systematic review

ronments [21]. AR has greatly improved over the last few decades, and todayAR usage is built into smartphone apps such as AR Travel Guide [22]. Theconcept of AR technology has been proven to have been effectively applied tomobile devices [23]. Rohs and Gfeller proposed the use of portable devices andsmartphones rather than specialised hardware to build AR apps [24]. Accordingto Hollerer and Feiner, mobile AR is a combination of various components suchas display technology, computational platform, global tracking technologies, dataaccess technology, wireless communication and interaction technology [25].


There are many apps with AR features developed for various purposes. Thesefeatures have been found to enhance the mobile applications’ usability. There weredifferent approaches discussed by experts to improve learnability of mobile apps[26]:

• Improving graphical icon characteristics that present icon usability and concrete-ness

• Introducing multilayered interfaces to allow the users to adopt an improvedmental framework and reduce complexity

• Enhancing the mobile device’s interface through the use of a larger display thatallows for guidance and feedback in real time

Augmented mobile apps better enable the use of mobile devices amongst thosewith declining cognitive ability such as the elderly [27]. Kim and Dey discoveredthat the use of AR in the windshield displays of vehicles that help older peoplein cognitive mapping was very effective [28]. The findings showed a noticeabledecrease in terms of distractions and errors in navigation when compared to earliermodels of such windshield displays designed for the elderly. AR uses have alsobeen discussed in the tourism sector. A research identified the benefits of using ARin tourism mobile apps through developing and evaluating a tourist mobile app withaugmented reality. The results proved that AR enhances the tourist experience in aninnovative way. Therefore, AR apps in different industries can improve the qualityof service [29].

AR in mobile apps represents a great opportunity for better access to digital andprint library collections. Mobile apps with AR technology deliver an interactiveand engaging information experience. AR can help apps overlay graphical data, andthis makes apps with AR technology well-suited for engagement in both libraryand real-world off-site interaction with the content. A research by Hahn introducesmobile AR apps for next-generation library services and uses [30]. The study showsthat mobile AR apps can help augment browsing of physical book stacks, opticalcharacter recognition, facial recognition and library navigation. The paper suggestsmobile AR uses and apps in library settings as well as introduces a model todemonstrate a prototype interface.

3.2 Virtual Reality in Mobile Apps

VR is defined as a computer-simulated or immersive multimedia reality [31].VR technology uses computers to replicate an imagined or real environment. Itallows user interaction through simulating the user’s physical environment andpresence. VR can artificially create sensory experience, such as touch, sight, smelland hearing. The origins of VR came from the science fiction world. In 1935Stanley G. Weinbaum’s short story ‘Pygmalion’s Spectacles’ is considered as apioneer work of fiction that introduces VR. The story describes a VR systemoperated via goggles which used holographic simulations to record users’ fictional


experiences and incorporated the senses of smell and touch [32]. Bob Sproulland Ivan Sutherland in 1968 created the first AR and VR head-mounted display(HMD) system [33]. In 1978, MIT created the Aspen Movie Map that is one of themore famous hypermedia and virtual reality systems. This programme was moreof Aspen’s virtual simulation. People could explore a town’s streets in a ‘polygon’mode as well as two others labelled ‘winter’ and ‘summer’; two of these reliedon photographs. The developers’ purpose was to capture every possible journeythrough the city’s network of roads and streets. The third mode was a 3-D modelof the town [34]. Street View was then introduced by Google in 2007; it consistsof panoramic views of numerous worldwide locations that include indoor buildings,roads and rural areas. In 2010, a stereoscopic mode was introduced [35]. Virtualreality is used then in many mobile apps in order to enhance UX.

Mobile apps related to health, education and gaming now increasingly havevirtual reality features to increase the usability of the app. A research designedan educational game with virtual reality, and the results demonstrated that thegame was likeable and usable. The researchers, however, believed there was amplescope for improvement in likeability and usability to maximise educational benefits[36]. A study presented an environment in a demo that enabled users to exploredifferent three-dimensional (3D) visualisations on tablets and smartphones [37]. Aperformance- and feedback-based app was tested and compared to a gamed-basedone with virtual reality [38]. The aim was to examine their effects on aspects ofimmediate response to an exercise bout. The participants reported the app withvirtual reality had a more associative attentional focus.

VR offers a lot of useful apps for tourism and deserve greater attention fromtourism professionals and researchers. A study shows that with the continuousevolution of VR technology, the significance and number of such apps will increase.Marketing, planning and management, entertainment, heritage preservation, educa-tion and accessibility are some areas of tourism in which VR could prove to be veryvaluable [39]. The study also emphasises that new challenges and questions willemerge with further integration of VR and tourism. Tozsa discusses how VR canbe useful in public administration services. The article suggests that virtual realitymobile apps can help in the field of e-government and the services provided throughsuch apps have the simplest tools for navigation and a more attractive outlay thantraditional e-government websites. He believes that with future developments in VRsuch as 3-D, administration could have a variety of useful apps [40].

4 Evaluation and Future Work

Both AR and VR technologies have been applied in several mobile apps. They havebeen used in different fields, for example, education, transportation and tourism.Moreover, both technologies can be used amongst those with declining cognitiveability such as the elderly. As both technologies have proven that they can be usedsuccessfully in different fields by a various group of people, our future work will


involve developing fitness mobile apps that include the use of both AR and VRtechnologies as a motivational tool. The app will be designed for obese individualsto help them be motivated and lose weight to have a healthy lifestyle. The app willalso consider the four main motivational features from the recent survey.

5 Conclusion

The literature review shows that obesity is a major problem all over the world.Obesity is defined as excessive fat in the human body. The percentage of peoplesuffering from obesity is increased to around 15%. Obesity is one of the reasonsfor several chronic diseases such as diabetes. Several experts believe that obesitycan be fought and stopped by engaging obese individuals to perform in physicalactivities. However, it is hard to motivate or keep obese individuals motivated toperform physical activities for losing weight to have a better lifestyle. Yet, themajority of experts concluded that behaviour intervention could be the solution forchanging behaviour. The use of fitness mobile apps is becoming popular around theworld as 16% of smartphone owners use fitness and health apps. Several expertsclaim that fitness behaviour interventions can be gained via using fitness mobileapps especially for those people who are suffering from obesity.

Fitness mobile apps have unique features that play an important role in orderto motivate or keep obese individuals to do physical activates. According to theresults from a recent study, there are four key features that help to motivateobese individuals. These features are (1) goal settings; (2) monitoring, tracking andfeedback; (3) reminders and alerts; and (4) reward or gamification. However, thereare more new technologies (AR and VR) that become popular amongst mobile appusers and can be applied to help obese individuals to lose weight.

This study aims to investigate on the benefit of using both AR and VR in mobileapps. A systematic review of the most recent researches and articles that studiedthe use of AG and VR in mobile apps has been done. We found that AR andVR have played a major role to enable a better use of mobile devices and apps.They have been used in a variety of fields in mobile apps, for example, education,transportation and tourism. Moreover, both technologies are used by users with adeclining cognitive ability such as the elderly. These results have encouraged us toconsider applying both AR and VR when we start developing a new fitness mobileapp. Our assumption is that as both technologies have advantages in other fields,they can have a positive effect on the fitness field. The app will be designed anddeveloped specifically for people who are suffering from obesity and want to bemotivated to lose weight and have a better lifestyle.


References

1. Simopoulos, A. P., & van Itallie, T. B. (1984). Body weight, health, and longevity. Annals ofInternal Medicine, 100(2), 285–295.

2. Organization, W. H. (2016). Obesity and overweight. Retrieved 2 Oct 2016, from http://www.who.int/mediacentre/factsheets/fs311/en/

3. Fontaine, K. R., Redden, D. T., Wang, C., Westfall, A. O., & Allison, D. B. (2003). Years oflife lost due to obesity. JAMA, 289(2), 187–193.

4. Calle, E. E., Thun, M. J., Petrelli, J. M., Rodriguez, C., & Heath, C. W., Jr. (1999). Body-massindex and mortality in a prospective cohort of us adults. New England Journal of Medicine,341(15), 1097–1105.

5. Summerbell, C., Waters, E., Edmunds, L., Kelly, S., Brown, T., & Campbell, K. (2005).Interventions for preventing obesity in children (review). Cochrane Library, 3, 1–71.

6. Saris, W., Blair, S., Van Baak, M., Eaton, S., Davies, P., Di Pietro, L., Fogelholm, M., Rissanen,A., Schoeller, D., & Swinburn, B. (2003). How much physical activity is enough to preventunhealthy weight gain? Outcome of the Iaso 1st stock conference and consensus statement.Obesity Reviews, 4(2), 101–114.

7. Anderson, J. L., Antman, E. M., Bailey, S. R., Bates, E. R., Blankenship, J. C., Casey, D. E.,Jr., Green, L. A., Hochman, J. S., Jacobs, A. K., & Krumholz, H. M. (2009). Aha scientificstatement. Circulation, 120, 2271–2306.

8. Hill, J. O., & Wyatt, H. R. (2005). Role of physical activity in preventing and treating obesity.Journal of Applied Physiology, 99(2), 765–770.

9. Foster, G. D., Makris, A. P., & Bailer, B. A. (2005). Behavioral treatment of obesity. TheAmerican Journal of Clinical Nutrition, 82(1), 230S–235S.

10. Wadden, T. A., & Stunkard, A. J. (2002). Handbook of obesity treatment. New York: GuilfordPress.

11. Statista. (2014). Facts and statistics on Wearable Technology. https://www.statista.com/topics/1556/wearable-technology/

12. Pai, A. (2014). Nielsen: 46 million people used fitness apps in January. http://www.mobihealthnews.com/32183/nielsen-46-million-people-used-fitness-apps-in-january

13. Statista. (2016). Share of internet users who use health and fitness apps every month as of 3rdquarter 2015. https://www.statista.com/statistics/502195/health-and-fitness-app-access/

14. Khalaf, S. (2014). Health and fitness apps finally take off, fueled by fitness fanatics. http://flurrymobile.tumblr.com/post/115192181465/health-and-fitness-apps-finally-take-off-fueled

15. M. Essany. (2013). Mobile Health Care Apps Growing Fast in Number. http://mhealthwatch.com/mobile-health-care-apps-growing-fast-in-number-20052/

16. Fogg, B. J. (2002). Persuasive technology: Using computers to change what we think and do.Ubiquity, 2002(December), 5.

17. Alturki, R. M., & Gay, V. (2016). A systematic review on what features should be supported byfitness apps and wearables to help users overcome obesity. International Journal of Researchin Engineering and Technology, 5, 9.

18. Smith, C. (2017). 80 incredible Pokemon go statistics and facts (April 2017). http://expandedramblings.com/index.php/pokemon-go-statistics/

19. Carmigniani, J., Furht, B., Anisetti, M., Ceravolo, P., Damiani, E., & Ivkovic, M. (2011).Augmented reality technologies, systems and applications. Multimedia Tools and Applications,51(1), 341–377.

20. Azuma, R., Baillot, Y., Behringer, R., Feiner, S., Julier, S., & MacIntyre, B. (2001). Recentadvances in augmented reality. Computer Graphics and Applications, IEEE, 21(6), 34–47.

21. Milgram, P., & Kishino, F. (1994). A taxonomy of mixed reality visual displays. IEICETransactions on Information and Systems, 77(12), 1321–1329.

22. Nazri, N. I. A. M., & Rambli, D. R. A. (2014). Current limitations and opportunities inmobile augmented reality applications, Computer and Information Sciences (ICCOINS), 2014International Conference on: IEEE, pp. 1–4.


https://www.statista.com/topics/1556/wearable-technology/

http://www.mobihealthnews.com/32183/nielsen-46-million-people-used-fitness-apps-in-january

https://www.statista.com/statistics/502195/health-and-fitness-app-access/

http://flurrymobile.tumblr.com/post/115192181465/health-and-fitness-apps-finally-take-off-fueled

http://mhealthwatch.com/mobile-health-care-apps-growing-fast-in-number-20052/

http://expandedramblings.com/index.php/pokemon-go-statistics/


23. Adhani, N. I., & Awang, R. D. R. (2012). A survey of mobile augmented reality applications,1st international conference on future trends in computing and communication technologies:Citeseer, pp. 89–96.

24. Rohs, M., & Gfeller, B. (2004). Using camera-equipped mobile phones for interacting withreal-world objects. In A. Ferscha & I. C. O. Perv (Eds.), Advances in pervasive computing: Acollection of contributions presented at PERVASIVE 2004 (1st ed., pp. 265–271). Ann Arbor:Österr. Computer-Ges.

25. Höllerer, T., & Feiner, S. (2004). Mobile augmented reality, Telegeoinformatics: Location-based computing and services (p. 21). London: Taylor and Francis Books Ltd.

26. Leung, R., Findlater, L., McGrenere, J., Graf, P., & Yang, J. (2010). Multi-layered interfacesto improve older adults’ initial learnability of mobile applications. ACM Transactions onAccessible Computing (TACCESS), 3(1), 1.

27. Zhou, S., Chen, Z., Liu, X., & Tang, H. (2011). An “elder mode” of new generation phoneusing augment reality. Procedia Environmental Sciences, 10, 936–942.

28. Kim, S., & Dey, A. K. (2009). Simulated augmented reality windshield display as a cognitivemapping aid for elder driver navigation, Proceedings of the SIGCHI Conference on HumanFactors in Computing Systems: ACeM, pp. 133–142.

29. de la Nube Aguirre Brito, C. (2015). Augmented Reality Applied in Tourism MobileApplications, eDemocracy & eGovernment (ICEDEG), 2015 second international conferenceon: IEEE, pp. 120–125.

30. Hahn, J. (2012). Mobile augmented reality applications for library services. New LibraryWorld, 113(9/10), 429–438.

31. Burdea, G. C., & Coiffet, P. (2003). Virtual reality technology. Hoboken, New Jersey: Wiley.32. Weinbaum, S. G. (2015). Pygmalion’s spectacles. Auckland: The Floating Press.33. Norman’s J. (2017). Ivan Sutherland and Bob Sproull Create the First Virtual Reality Head

Mounted Display System. http://www.historyofinformation.com/expanded.php?id=108734. Lippman, A. (1978). The Aspen movie map. Cambridge: MIT ARPA.35. Lardinois, F. (2010). Google Street View in 3D: More Than Just an April Fool’s Joke. http://

readwrite.com/2010/04/06/google_street_view_in_3d_here_to_stay/36. Virvou, M., & Katsionis, G. (2008). On the usability and likeability of virtual reality games for

education: The case of Vr-engage. Computers & Education, 50(1), 154–178.37. Hürst, W., Beurskens, J., & van Laar, M. 2013. An experimentation environment for mobile

3d and virtual reality, Proceedings of the 15th international conference on Human-computerinteraction with mobile devices and services: ACM, pp. 444–447.

38. Gillman, A. S., & Bryan, A. D. (2015). Effects of performance versus game-based mobileapplications on response to exercise. Annals of Behavioral Medicine, 50, 1–6.

39. Guttentag, D. A. (2010). Virtual reality: Applications and implications for tourism. TourismManagement, 31(5), 637–651.

40. Tozsa, I. (2013). Virtual reality and public administration. Transylvanian Review of Adminis-trative Sciences, 9(38), 202–212.

http://www.historyofinformation.com/expanded.php?id=1087

http://readwrite.com/2010/04/06/google_street_view_in_3d_here_to_stay/

Cloud-Assisted IoT-Based SmartRespiratory Monitoring Systemfor Asthma Patients

Syed Tauhid Ullah Shah, Faizan Badshah, Faheem Dad, Nouman Amin,and Mian Ahmad Jan

1 Introduction

The weight of the conventional healthcare system is becoming substantial becauseof the increasing number of aging people combined with the small amount ofhealthcare resources and personnel (doctors, hospital administrators, nurses). In thehealthcare industry, the use of communication apps, things (sensors and devices)for healthcare monitoring are increasing exponentially and have a huge impact onthe patients and healthcare professionals. According to Forbes and Gartner, it isestimated that by 2020, approximately 20 billion smart devices will be connected tothe Internet [1], while HealthcareIoT market will contribute to $117 billion with anannual increasing rate of 38% during the 6-year period from 2016 to 2022 [2, 3].According to these assessments, it is predicted that HealthcareIoT can play a centralrole in the healthcare industry. IoT has a significant effect on the healthcare industry,and various technologies, tools, and devices are used for different types of patients(e.g., temperature, pulse, and oxygen in the blood, blood pressure, glucose monitor,GSR, and ECG) to minimize avoidable deaths.

Presently, HealthcareIoT is in its early phases with respect to deployment, design,and development. Nevertheless, IoT-based emerging technologies are providinga significant influence, and soon many healthcare monitoring technologies willemerge. By minimizing preventable deaths that happen due to hospital errors,HealthcareIoT has the potential to save more than 50,000 patients in the UnitedStates every year [4]. It guarantees patient safety and welfare by collecting andmonitoring patient health information with related support resources (wearablesmart devices to collect data and facilities for healthcare staff). HealthcareIoT can

S. T. U. Shah · F. Badshah · F. Dad · N. Amin · M. A. Jan (�)Department of Computer Science, Abdul Wali Khan University Mardan, Mardan, KPK, Pakistane-mail: [email protected]


77



https://doi.org/10.1007/978-3-319-96139-2_8

78 S. T. U. Shah et al.

Network Recourses Decryption

Networkproviders

Researchinstitution Emergency aid

services

Dispensary anddrugmanufacturer

Family,friendsand social media

HealthCareadvisor

Patient withsensor anddevices

Analysis StorageProcessing servers

Fig. 1 Conceptual design of the proposed HealthcareIoT system

result in access to better healthcare, minimized cost, and direct interaction betweenhealthcare professionals and the patient (Fig. 1).

Asthma is a prolonged lung disease of wheezing and breathlessness that occursseveral times a day or in a week in individuals. This disease is common amongchildren, and the number of deaths in children is greater than those in the elders.According to WHO, more than 235 million people are suffering from asthma [5].This disease occurs in both low-income and high-income countries, while more than80% of deaths occur in low-income countries. According to estimates released in2016 by WHO, over 383,000 deaths occur because of asthma in 2015. Air pollutionand smoking are the core causes for the problem. Due to the absence of instanthelp, elder people suffer most and may experience stress. Still, no widespread workhas been done about cloud-assisted HealthcareIoT-driven framework for asthmapatients.

In this paper, we proposed a cloud-assisted IoT-driven healthcare monitoringframework for asthma patients. First of all, using different devices and sensor,patient respiration rate can be recorded at home or outdoors. On the client side,using a desktop based or mobile application will take out the undesirable noisefrom the collected signal. As patient data safety is very important in HealthcareIoT.Therefore, data must be protected from unauthorized access. A gap in the patientdata security may cause mental disorder, social embarrassment, heart attack, or anyother physical damage. Therefore the signal is embedded with watermarking for

Cloud-Assisted IoT-Based Smart Respiratory Monitoring System for Asthma Patients 79

authentication and security. Respiration rate is calculated with the help of a simplealgorithm. After that the encrypted signal is transmitted to the cloud, where featuresare extracted and classified with the help of a classifier. Then the classified datawith watermarked signal is transmitted to the specific healthcare professional. Thehealthcare professional analyzes the signal and classified data and sends a decisionto the cloud server. At the end, the patient is notified with the healthcare professionaldecision.

2 Literature Review

The IoT is an emerging technology of interconnected devices and sensors. Thesedevices and sensor can capture, store, transmit, and share that data for classificationand analysis. Many innovative IoT applications emerge in the decade [6–18]. But themost promising application between them is healthcare monitoring. HealthcareIoTconsists of collecting patient data through different devices and sensors, analyzing,and transmitting them by using a network to healthcare advisors for immediate care[6–8, 12, 13, 15].

Respiratory rate is one of the important physiological aspects to be monitoredto obtain patient health information in critical condition. It is an important signto predict illness, a rapid refuse in patient health. In [16] the authors proposedweb-based gateways for IoT eHealthcare monitoring with both wireless and wired-based services. The wired gateways are used in a small building or room to makethe system low cost and power efficient; movement inside the building or roomis restricted. In [17, 18] radio-frequency identification (RFID)-based healthcaremonitoring solutions were proposed. In [18] an IoT-based framework was proposedto monitor patient data by using UHF-RFID technology. In [17] author proposed ahealth monitoring system using RFID to capture patient humidity and temperatureinformation and send them to the cloud for further analysis and understanding. In[19] an asthma monitoring scheme was developed using vital signs and IoT. In [6]authors investigated various challenges and prospects exist in IoT for managementand health monitoring. In [7] using cloud computing, mobile computing, and webservices, the authors developed a remote patient monitoring system.

In [14] a secure data collection and transmission, mobile-based healthcarearchitecture was introduced. Secure patient information is collected with the cryp-tography, private, and secret keys. The data is securely transmitted with authenticuser access by using attribute-based encryption scheme. The core disadvantageof this scheme is the computation time. In [19] a web-enabled IoT-based patientrespiration data collection scheme was introduced using vital sign. The mainproblem with this approach is secure communication.

Till now, no secure comprehensive work was found in cloud-assisted IoT-basedhealthcare monitoring for asthma patients. Where the respiration signal is collectedusing sensors, the collected signal is enhanced and watermarked before transmittingto the cloud server using the Internet. At the cloud end, features are extracted afterdecryption and signal restoration, where the healthcare professional analyzes patientcondition and provides immediate care.


3 Proposed Cloud-Assisted IoT-Based Health MonitoringSystem

IoT-based healthcare monitoring can play a very important role in modern-daypatient care by adapting a wide range of interconnected sensors and devices;cloud technology and big data are used to collect, record, and monitor patientdata. The integration of these technologies can form a smart healthcare patientmonitoring system. The proposed scheme can transfer patient information amongthe different participants in a secure manner, where it is only available to theauthorized healthcare monitoring team. At the cloud side, it supports to analyze,store, monitor, and securely transmit the patient data for medical recommendationsand additional evaluation to minimize hospital errors and provide quality healthcare.Figure 2 illustrates the IoT-based health monitoring scheme for asthma patients.First, patient respiration data are collected using a sensor, and then after signalrestoration, enhancement, and watermarking (encryption), it is transmitted using anetwork connection to the cloud server.

Fig. 2 The proposed scheme for patient respiration monitoring


The key components of the proposed scheme are described as follows.

Respiration Signal Collection Service This service is used to capture and storepatient respiration signal from various sensors and devices.

Healthcare Advisor and Other Support Services Patient respiration signal isuploaded to the cloud server, where it is stored in the cloud database, so healthcareadvisors can access and analyze that signal for possible action and medication.

Recourse Allocation Management This phase is responsible for allocation ofphysical resources and regulating VMs. It also provides authorized data access tothe patient information and decides whether to share the information or not.

Secure Signal Transmission Service This service allows secure transmission ofthe record respiration signal. This can be achieved by applying watermarking to therecorded signal.

Cloud Manager This service manages all the recourse require by each service andalso regulator each VMs, i.e., respiration signal collection and storage management,signal enhancement and restoration, information extraction and classification andlastly information acquisition.

1. Signal collection and storage management: this phase provides web-basedservices to manipulate and store patient health information in the data.

2. Signal enhancement and restoration: this service is responsible for enhancing thequality of decrypted signal.

3. Signal classification and analysis: this service is responsible for tracking andmonitoring activates.

IoT-Added Service This service records respiration signal and stores it. It providescontinuous monitoring by collecting patient respiration data continuously in bothoutdoor and home conditions and securely transmitting them to communicationgateway. These services include signal capturing, feature extraction and classifica-tion, and secure transformation. The healthcare advisor can access the data withouthaving any direct interaction with patients.

4 Proposed Health Monitoring Scheme

4.1 Signal Enhancement

To get rid of common glitches and noise from the recorded signal, it must beenhanced before any kind of processing and feature extraction. In the enhancementphase, low-pass filter is applied to the recorded signal to overwhelm high frequencymodules caused by noise. After that, the median filter was applied to eliminateminute glitches. Figure 3 shows the effect of passing the signal through the low-passand median filters. The figure illustrates that the quality of the signal is improvingand looks clear after applying the low-pass and median filters.


Fig. 3 The effect of low-pass and median filter on record signal: (a) recorded signal, (b) low-passfiltered, and (c) median filter signal

4.2 Watermarking

This technique is used to protect signals from forgery. Through this technique, weembed some specific information in the signal so that its credibility can be protectedand the signal will no longer intermix with other signals. This technique consists oftwo parts which are watermark embedding and watermark extraction. Watermarkembedding adds some information to the respiration signal and guarantees theauthenticity. We used discrete wavelet transform (DWT) for this purpose. Thistechnique decomposes the respiration signal to four different frequencies of sub-bands: horizontal details (HLi), low-frequency details (Li), high-frequency details(HHi), and vertical details (LHi). We apply the watermark embedding into thosecoefficients having low frequencies [20]. We used IDWT technique to extract theoriginal details of the signal [21]. Watermark embedding to the signal is applied onthe client, while extraction is applied at cloud side.



At the cloud server side, different features are extracted from the respiration signal.That includes calculating the breath rate/minute (bpm). Breath rate may varybetween 12 and 37 bpm, while the normal breath rate of an adult in resting positionis 12–20 bpm. The breath rate for a 1-year-old baby can vary between 30 and 40,while in elders, it may vary between 12 and 30. To calculate the breath rate, we usedthe algorithm in Fig. 4.

Algorithm_1 – CalculateRespiratoryRate

Input: Take numerical data of 20 secondsOutput: Breath Rate/Minute (bpm)Place the data in resp[]Numofpeaks = 0for i = 0 to resp.length - 5

if(resp [i]< resp [i+2] && resp [i+1]< resp [i+2] &&resp [i+3]> resp [i+2] && resp [i+4]> resp [i+2] )

Thennumofpeaks + 1

enf ifend forbpm = numofpeaks x 3

In the algorithm described in Fig. 4, first of all, we will take data of 20 s fromthe graph. In the next step, the data will be placed in an array. Next, using a slidingwindow technique with having a size of 5 for each window, we identify the totalnumber of peak points. Through this we will compute each set of five elements; if

Fig. 4 Classification accuracy of the proposed scheme


the central element is greater than the central point, then it will be considered aspeak point, and the numofpeak variable will be incremented. After computing allthe data in the array, the breath per minute is obtained by multiplying the resultantnumofpeak by 3.

5 Experimental Results and Evaluation

5.1 Classification Performance

For classification purposes, we performed two different kinds of experiments: onewith the data captured though the proposed approach and the other with data collectfrom MIMIC II Waveform Database [22]. The experiential results and accuracyrates of both databases were shown in Fig. 5. The accuracy was evaluated withdifferent features. With 40 features the accuracy rate reached to 87% with privatedatabase and 83% with the MIMIC II database.

Fig. 5 Workload for the transmission of respiration signal


5.2 Workload of the Proposed System

Figure 5 shows the workload of the transmission services used by the proposedsystem. The figure focuses on the transmission of respiration signal. The runtimemeasurement of this service is collected using a Java-based simulator program. Toachieve this, we used a virtual machine with Intel core i3 1.7 GHZ processor, 3GBDDR ECC RAM, 8MBPS bandwidth, and running windows server 2012.

6 Conclusion

IoT-based health monitoring is an emerging technology that can revolutionize thehealth industry. In the hospital-centric healthcare service, patient health status isrecorded through different processes and devices. In this chapter, we developeda smart, inexpensive, scalable, and efficient IoT-based cloud-assisted healthcaremonitoring framework for asthma patients. The proposed scheme provides remotemonitoring of patient health status anytime, anywhere, and enables healthcareadvisors to access, analyze, track, and monitor patient health status in a real-timemanner. To ensure security, the recorded signal was watermarked before sending tothe cloud server. The performance measurements and classification accuracy wereevaluated through experiments. Future work involves implementing the proposedsystem with real patients and healthcare advisors as a test trial.

References

1. https://www.forbes.com/sites/tjmccue/2015/04/22/117-billion-market-for-internet-of-things-in-healthcare-by-2020/#5a6eba4069d9

2. https://www.psmarketresearch.com/market-analysis/smart-home-healthcare-market3. http://healthitanalytics.com/news/healthcare-internet-of-things-driving-global-market-growth4. http://blog.iiconsortium.org/2015/01/how-toindustrial-internet-of-things-can-save-50000-

lives-a-year.html5. http://www.who.int/mediacentre/factsheets/fs307/en/6. Hassanalieragh, M., et al. Health monitoring and management using Internet-of-Things (IoT)

sensing with cloud-based processing: Opportunities and challenges, Services Computing(SCC), 2015 IEEE international conference on. IEEE, 2015.

7. Mohammed, J., et al. Internet of Things: Remote patient monitoring using web services andcloud computing. Internet of Things (iThings), 2014 IEEE international conference on, andgreen computing and communications (GreenCom), IEEE and cyber, physical and socialcomputing (CPSCom), IEEE. IEEE, 2014.

8. Hu, L., et al. (2015). Software defined healthcare networks. IEEE Wireless Communications,22(6), 67–75.

9. Riazul Islam, S. M., et al. (2015). The internet of things for health care: A comprehensivesurvey. IEEE Access, 3, 678–708.

https://www.forbes.com/sites/tjmccue/2015/04/22/117-billion-market-for-internet-of-things-in-healthcare-by-2020/#5a6eba4069d9

https://www.psmarketresearch.com/market-analysis/smart-home-healthcare-market

http://healthitanalytics.com/news/healthcare-internet-of-things-driving-global-market-growth

http://blog.iiconsortium.org/2015/01/how-toindustrial-internet-of-things-can-save-50000-lives-a-year.html



10. Jara, A. J., Zamora-Izquierdo, M. A., & Skarmeta, A. F. (2013). Interconnection frameworkfor mHealth and remote monitoring based on the internet of things. IEEE Journal on SelectedAreas in Communications, 31(9), 47–65.

11. Xu, B., et al. (2014). Ubiquitous data accessing method in IoT-based information system foremergency medical services. IEEE Transactions on Industrial Informatics, 10(2), 1578–1586.

12. Li, Y., Guo, L., & Guo, Y. Enabling health monitoring as a service in the cloud. Utility andcloud computing (UCC), 2014 IEEE/ACM 7th international conference on. IEEE, 2014.

13. Hossain, M. S. (2015). Cloud-supported cyber–physical localization framework for patientsmonitoring. IEEE Systems Journal, 11(1), 118–127.

14. Zhang, K., et al. (2015). Security and privacy for mobile healthcare networks: From a qualityof protection perspective. IEEE Wireless Communications, 22(4), 104–112.

15. Hossain, M. S., & Muhammad, G. (2014). Cloud-based collaborative media service frameworkfor healthcare. International Journal of Distributed Sensor Networks, 10(3), 858712.

16. Granados, J., et al. (2015). Web-enabled intelligent gateways for eHealth internet-of-things. InInternet of Things. User-Centric IoT (pp. 248–254). Cham: Springer International Publishing.

17. Amendola, S., et al. (2014). RFID technology for IoT-based personal healthcare in smartspaces. IEEE Internet of Things Journal, 1(2), 144–152.

18. Catarinucci, L., et al. (2015). An IoT-aware architecture for smart healthcare systems. IEEEInternet of Things Journal, 2(6), 515–526.

19. Raji, A., et al. Respiratory monitoring system for asthma patients based on IoT. Greenengineering and technologies (IC-GET), 2016 online international conference on. IEEE, 2016.

20. Tiwari, N, Ramaiya, M. K., & Sharma, M. Digital watermarking using DWT and DES.Advance computing conference (IACC), 2013 IEEE 3rd international. IEEE, 2013.

21. Gonge, S. S., & Bakal, J. W. (2013). Robust digital watermarking techniques by using DCT andspread spectrum. International Journal of Electrical, Electronics and Data Communication,1(2), 111–124.

22. Goldberger, A. L., et al. (2000). Physiobank, physiotoolkit, and physionet. Circulation,101(23), e215–e220.

Blood Cell Counting and SegmentationUsing Image Processing Techniques

Ayesha Hoor Chaudhary, Javeria Ikhlaq, Muhammad Aksam Iftikhar,and Maham Alvi

1 Introduction

The blood is the red thick aqueous substance that runs through the entire human oranimal body, via arteries and veins. According to the biologists, the blood amountsup to 7% of the human body weight, approximately [1]. The main functions of theblood traveling through our bodies are to provide oxygen to the body organs, to takeback the carbon dioxide from the organs to the respiratory system, to fight againstthe infectious substances, and also to provide nutrients to all the body organs. Bloodis essential for the survival of the vertebrates and some of the invertebrates as well.A complete blood count plays a vital role in determining a person’s health rate.Doctors perform a complete blood count in order to obtain important informationsuch as kinds and numbers of cells in the blood. This information can further guidethem in the diagnosis or screening of various health issues.

The blood constitutes four elements that are as follows (see Fig. 1):

Plasma, a thick aqueous solution carrying blood cells which provide nutrients to thebody and get rid of the waste materials.

Next are the red blood cells (or erythrocytes). Hemoglobin is the substance thatgives blood its color and is a constituent of RBCs. These cells are important forthe transportation of oxygen to the body.

A. H. Chaudhary · J. Ikhlaq · M. A. Iftikhar (�)Department of Computer Science, COMSATS Institute of Information Technology, Lahore,Pakistan

M. AlviPunjab University College of Information Technology, Quaid-e-Azam Campus, University of thePunjab, Lahore, Pakistan


87


https://doi.org/10.1007/978-3-319-96139-2_9

88 A. H. Chaudhary et al.

(a) (b)

(c) (d)

PlasmaBloodvessel

Fig. 1 Images of (a) blood plasma, (b) red blood cells or erythrocytes, (c) white blood cell orleukocyte (encircled), and (d) platelets or thrombocytes (encircled)

Succeeding now are the white blood cells or leukocytes; these are called thedefender cells as they fight against the germs that cause infections.

Lastly, platelets or thrombocytes perform the process of clotting in case of bloodvessel rupture.

It is important to mention here that red blood cells are present in a large amountthan the rest of the cells. Deficiency or abundance of any of these three types ofblood cells can result in various health issues such as anemia [2], leukemia [3],or sickle cell disease [4]. To prevent from such diseases, it is vital to determinean accurate blood count. For that purpose, various kinds of blood tests can beperformed such as blood culture, bone marrow biopsy, and Coombs test, but all ofthese are specialized tests performed under special circumstances. Complete bloodcount (CBC), on the other hand, is the type of blood test for a regular check-up.

CBC is an automated test of the blood to analyze and determine the numberand kinds of cells present in the human blood, thus indicating any changes such asabundance or deficiency of any of the blood cells. There are two popular methodsto perform a complete blood count, manually and by the help of automated bloodanalyzer. But there are also some major shortcomings to both these methods.Performing CBC manually can be a very extensive task, especially when bloodsample to be analyzed is large. Automated but expensive analyzers provide goodresults, but under normal conditions [5], they do not identify abnormal blood. So,the problem in both methods leads to inaccuracy of the results.

Blood Cell Counting and Segmentation Using Image Processing Techniques 89

In literature, extensive work has been proposed by researchers for blood cellsdetection and counting using image segmentation techniques. For example, oneapproach [6] introduced Hough transform technique to estimate the number of redblood cells. Similarly, another [7] presented a segmentation approach that couldbe used to remove white blood cells and platelets from blood smear images. Aresearch conducted by Shashi Bala and Amit Doegar [8] presented the segmentationand identification of red blood cells and white blood cells through color-basedsegmentation using CIE L*a*b* (CIELAB). Another algorithm suggested animage contouring technique that utilized scanning electron microscopic images [9].Another approach was suggested to identify malaria parasites by counting the redblood cells in the blood smear [10]. Their proposed methodology consisted of twoparts. The first one dealt with the identification of malaria parasite (i.e., the blue staindeposits) utilizing Zack’s method [11] in order to calculate the threshold value of theS component image, extracted from HSV color space. Heidi Berge [12] combinedthree methods to count the red blood cells in the blood smear images. The firsttechnique is the application of morphological methods for red blood segmentation.Then thresholding is utilized for boundary curvature calculations. The third and lasttechnique applied was Delaunay triangulations [13] to split red blood cell clumps.

F. Sadeghian [14] proposed the pixel intensity thresholding in order to segmentthe cytoplasm in the white blood cells. This method completely separates thebackground of the image from its other components. Before applying intensitythresholding, they also applied Snake algorithm [15] to segment white blood cellsnucleoli of various shapes and colors. Laplacian of Gaussian (LoG) [16] edgedetector was also used, in order to produce an image which is not over-segmentedby detecting the edges of the image. This approach [17] suggested that watershedalgorithm segmented the images in an overly manner, and in order to avoid this,edge detection can be utilized. To predict cancer cells in blood samples, an approachbased solemnly on image processing techniques was proposed by Jagadeesh [18].This approach involved the segmentation of the image of the bone marrow aspiratewith the help of watershed algorithm. This approach involved the segmentation ofthe image of the bone marrow aspirate with the help of watershed algorithm, alongwith other few morphological operators such as opening and closing, erosion, anddilation. Wold’s decomposition model was introduced for the feature extraction ofthe blood cell images.

We will be comparing the proposed technique with two others working for thesame purpose to check the extent of accuracy and speed under the same conditions.The Red Blood Cell Segmentation Using Masking and Watershed Algorithm: APreliminary Study [7] provides a pragmatic approach to count the red blood cellsin a given image by subtracting the white blood cells and platelets from the imageand the application of watershed algorithm on the subtracted image. The process forthe removal of white blood cells and platelets involves the filtration of the image,color conversion using Ycbcr, binary erosion, the application of morphologicalreconstruction, and hole filling. The morphologically reconstructed image is thensegmented using Laplace of Gaussian (LoG) edge detection, gradient magnitude,and marker-controlled watershed algorithm.


Another study [20], suggests this method involves the utilization of circularHough transform (CHT). Preprocessing techniques such as thresholding and imagecontrast are applied to the input image in order to extract the RBC from the image.

Then, circle Hough is applied to the contrast image to analyze the RBC based onthe minimum and maximum radius of RBC.

Some bigger problems faced during the implementation of both these methodsare the cells that are on the corner of the image/blood smear slide and are oftennot counted. The overlapped cells are counted only once. The size of blood cellsespecially the red blood cells varies, therefore making it harder to count each cellindividually. The proposed technique aims to develop a system that overcome thesedrawbacks to generate accurate results, using image segmentation techniques.

The rest of the article is structured as follows: Section 2 describes the proposedmethod. Section 3 presents performance evaluation measures. Experimental resultsare described in Sect 4, and Sect. 5 concludes the paper.


In this article, a blood cell counting and segmentation system has been proposed,which uses feature-based segmentation, erosion, image subtraction, and convexhull and convex segmentation for detecting and eventually counting blood smearsamples. Employing the said image analysis techniques, an algorithm has beendevised to perform blood cell image analysis, segmentation, and cell counting. Thealgorithm has been divided into three stages: the first stage is the extraction ofwhite blood cells, and the other two stages involve the extraction of single cellsand overlapped cells (Figs. 2).

2.1 Extraction of the White Blood Cells

The first step involves segmentation of the blood smear image in l*a*b space.Segmentation of an image is a cardinal task in the process of image analysis,which constitutes of division of an image into multiple segments for enhancedidentification of objects on interest in contrast to their background. The image inl*a*b space is divided into three layers, namely, l, *a, and *b, where “l” stands forbrightness or luminosity, “*a” stands for the hue and saturation layer along the red-green axis, and “*b” stands for the hue and saturation layer along the blue-yellowaxis. Conversion of the image into l*a*b color space separates the blue channel,which results in segmentation of the nuclei in the image.

The first step of the algorithm is to acquire an image and then segmentation onthe “original image” based on l*a*b color space. Here, l, *a, and *b represent threedifferent layers of the image, where “l” stands for brightness or luminosity, “*a”stands for the hue and saturation layer which shows the colors that fall along the


Start

ImageAcquisition

ImageAcquisition

Total RBC’s

RBC−Phase IIRBC−Phase I

Total Blood Count

Total WBC’s

WBC − White Blood Count

K- MeanClustering After WBC Completes

After RBC − Phase I Completes

Hole Filling

FeatureExtraction

Segmentati-on

NucleusExtraction

Masking &Reconstruct

-ion

Filling ofthe Nucleus

Convex Hulland

Deficiency

Convex Hulland

Deficiency

Concavity

ApplyingMedianFilter

Fig. 2 Schematic overview of the proposed methodology

red-green axis, and “*b” stands for the hue and saturation layer which shows thecolors that fall along the blue-yellow axis. Conversion of the image into l*a*b colorspace results in the separation of the blue channel, which in simple words can bedescribed as the separation of the nucleus/nuclei from the rest of the objects of theimage.

After segmentation, the image is then eroded in order to “filter” all the noisepresent in the image. “Image noise” is a term used to express the variations in theimage brightness or color information.

2.2 RBC Extraction Phase I

In this step the image is binarized. Binarization of the image means conversion of theimage into binary pixel values, i.e., 0 or 1, where 0 and 1 stand for black and white,


Fig. 3 Single Cells

respectively. One of the most conventional techniques to perform image binarizationis the K-means algorithm. Figure 5 is the image obtained after the application of K-means algorithm; we can see that this image is free of false connected components.

K-means Algorithm A brief explanation of all the steps involved in the K-meansalgorithm is given: The first step of the k-means algorithm is to abstractly findthe center of each cluster of objects present in the image. Place these centers ata distance to each other to get more filing results. Once each cluster of object hasbeen found, the next step is to calculate the distance between each data point andcluster. Now at the third step, the data point is assigned to the cluster center withminimum distance from all the cluster centers. Recalculate the new cluster centerusing:

vn = (1 cn)∑

xn cn m = 1 (1)

where “cn” represents the number of data point in nth cluster [21]. Recalculate thedistance between each data point and new obtained cluster centers. If no data pointwas reassigned, then stop; otherwise repeat k-means clustering algorithm, n.d.

The holes that can be seen in the cells in “converted image” (Fig. 6) are nowfilled using hole filling. In the figure we can see tiny cell-like objects. These objectsare the platelets; MATLAB function bwareaopen is used to remove the platelets.Threshold is set using hit and trial method.


2.3 RBC Extraction Phase II

In this step we discover the properties of the connected regions of the image. For thispurpose, we have to define a metric to test circularity. The most common measurefor circularity is perimeter and area. The metric for circularity can thus be given as(Shape Factor (Image Analysis and Microscopy), n.d.):

circularity = perimeter2/4πA (2I)

A single cell should have circularity near to 0.9, as 1 is the maximum value forcircular objects. After the implementation of the abovementioned metric, we havenow the values for perimeter, area, and centroid of the cells. These features will aidin the differentiation of the single cell from overlapped cells.

In this last step, we find the conventional object description, using convex hulland convex deficiency. Convex hull is used to find the convex deficiency of eachconnected component. It may be noted that, in general, the convex deficiencies(number of concave regions) corresponding to each connected component provideactual number of overlapping single cells in the connected component. Therefore,in simple words, we can say that convex hull discovers the hidden objects of theimage.

Convex hull and convex deficiency are two techniques that are functional incases where you need to get a conventional object description. The mathematicaldescription of a convex hull is stated as, a set A is said to be “convex” if the straightline segment joining any two points in A lies entirely within A. The convex hull “H”of an arbitrary set S is the smallest convex set containing S. The difference betweenH and S is called the convex deficiency [22].

3 Performance Evaluation Measures

This article includes a comparison of the proposed algorithm with the three othertechniques. The first technique is manual counting of blood cells in the image.The other two techniques are based on the two former research papers, Red BloodCell Segmentation Using Masking and Watershed Algorithm [7] and AutomatedRed Blood Cells Counting in Peripheral Blood Smear Image Using Circular HoughTransform [20].


Table 1 Result comparison

Image

Results usingproposedtechnique (R1)

Results obtained fromimplementation of J.M. Sharif, 2012 (R2)

Results obtained fromimplementation ofMazalan, 2013 (R3)

Results usingmanual counting(R4)

RBC Acc. RBC Acc. RBC Acc. RBCImage 1 884 84% 678 90% 1066 69% 746Image 6 568 98% 351 62% 957 58% 562Image 12 564 97% 963 60% 963 60% 579Image 18 884 88% 1448 51% 1044 71% 747Image 19 934 84% 1198 92% 2022 55% 1108Image 22 931 92% 1469 68% 2206 50% 1006Image 27 62 80% 700 77% 1007 53% 539Image 34 643 84% 1051 50% 809 65% 533

Accuracy percentage average of the proposed algorithm = 89%Accuracy percentage average of technique 1 = 68%Accuracy percentage average of technique 2 = 60%

3.1 Comparison of the Results

The results obtained by the implementation of the proposed algorithm and theresults acquired from manual counting and by the application of the abovementionedtechniques are given in Table 1.

The results obtained from the implementation of the proposed algorithm areefficient in case of RBCs, although some minor flaws are seen in case of whiteblood cells, which might be the reason why we utilized L*a*b color space. Theerrors in the results obtained by the technique are due to over-segmentation of thecells. It includes the radius of platelets along with other cells.

4 Experimental Results

To verify the authenticity of the algorithm, we performed it on other images. Thischapter includes test cases and their results.

1. Original Image (Fig. 4)2. Image after extracting nucleus then performing hole filling and erosion (Fig. 5)3. Converting the subtracted image into binary image using k-means algorithm

using (2) (Fig. 6)4. Image obtained after subtracting nucleas and filling the region with background,

then converting to binary (Fig. 7)5. Number of extracted overlapped cells calculated as ‘NumObjects’ (Fig. 8)


Fig. 4 Original image

Fig. 5 Image after extracting the nucleus, then performing hole filling and erosion


This paper presents a software tool to perform an automated analysis of the bloodsmear image using digital image processing techniques. The approach utilizes themorphological approaches for image segmentation, extraction, and estimation. Asper the methodology presented above, we have successfully separated the white


Fig. 6 Using (Fig. 5) as amask to mark which areas aregoing to be filled

Fig. 7 Image obtained aftersubtracting the nucleus, fillingthe region with background,and then converting to binary

Fig. 8 Number of extractedoverlapped cells calculated as“NumObjects”


blood cells from the other cells and the other objects of the image. Platelets havealso been removed using an area limit; thus the number of platelets can be countedby counting the objects separated. The number of overlapping cells present in theblood cell image has also been successfully identified and counted, using featuressuch as cell area, perimeter, and concavity.

Our research has improvements as contrary to the standard; local data has beenused, from Agha khan laboratories, and also we have tackled red blood cells, whiteblood cells, and platelets. Comprehensive work involving all three cell types was oneof the focal points in our paper. So our research gives a much simpler and cheapertechnique with accuracy better than the manual method.

The output of the current research is simply for the hematologists to conduct acomplete blood count test without putting in a lot of effort. This might also help inreducing the time taken to perform the CBC test calculations. Finally, it is useful fordetection of certain diseases, which are diagnosed based on blood cell count.

For further research we would use neural nets and machine learning to increasethe accuracy. We could also use our data as training set to predict whether the cellsin test data are normal or abnormal, which could be further used to detect the actualdisease of a person.

References

1. Kasuya, H., Onda, H., Yoneyama, T., Sasaki, T., & Hori, T. (2003). Bedside monitoring ofcirculating blood volume after subarachnoid hemorrhage stroke. Stroke, 34(4), 956–960.

2. Visconte, V., Tabarroki, A., Gerace, C., Al-Issa, K., Hsi, E. D., Silver, B. J., Lichtin, A. E.,& Tiu, R. V. (2014). Somatic mutations in splicing factor 3b, subunit 1 (SF3B1) are a usefulbiomarker differentiate between clonal and non-clonal causes of sideroblastic anemia. Blood,124, 5597.

3. Smith, M. T. (1996). The mechanism of benzene-induced leukemia: A hypothesis andspeculations on the causes of leukemia. Environmental Health Perspectives, 104(Suppl 6),S1219–S1225. https://doi.org/10.1289/ehp.961041219.

4. Audard, V., Bartolucci, P., & Stehlé, T. (2017). Sickle cell disease and albuminuria: recentadvances in our understanding of sickle cell nephropathy. Clinical Kidney Journal, 10(4), 475–478.

5. Aroon Kamath, M. (2014). Automated blood-cell analyzers. Can you count on them to countwell?. Doctors Lounge Website. Available at: https://www.doctorslounge.com/index.php/blogs/page/17172

6. Mahmood, N. H., & Mansor, M. A. (2012). Red blood cells estimation using Hough transformtechnique. Signal & Image Processing, 3(2), 53.

7. Sharif, J. M., Miswan, M. F., Ngadi, M. A., Salam, M. S. H., & Bin Abdul Jamil, M. M. (2012).Red blood cell segmentation using masking and watershed algorithm: A preliminary study. InBiomedical Engineering (ICoBE), 2012 International Conference on (pp. 258-262). IEEE.

8. Bala, S., & Doegar, A. (2015). Automatic detection of sickle cell in red blood cell usingwatershed segmentation. International Journal of Advanced Research in Computer andCommunication Engineering, 4(6), 488–491.

9. Joost Vromen, B. M. (2009). Red blood cell segmentation from SEM images. In: Imageand Vision Computing New Zealand, (2009). IVCNZ’09. 24th International Conference, NewZealand.

http://dx.doi.org/10.1289/ehp.961041219

https://www.doctorslounge.com/index.php/blogs/page/17172


10. Damahe, L. B., Krishna, R., & Janwe, N. (2011). Segmentation based approach to detectparasites and RBCs in blood cell images. International Journal of Computer Science andApplications, 4, 71–81.

11. Zack, G. W., et al. (1997). Automatic measurement of sister chromatid exchange frequency.The journal of histochemistry and cytochemistry: Official Journal of the HistochemistrySociety, 25(7), 741–753.

12. Scholz, J., Klein, M. C., Behrens, E. J., & Johansen-Berg, H. (2009). Training induces changesin white matter architecture. Nature Neuroscience, 12, 1370–1371.

13. Lee, D. T., & Schachter, B. J. (1980). Two algorithms for constructing a delaunay triangulation.International Journal of Computer and Information Sciences, 9(3), 219–224.

14. Sadeghian, F., Seman, Z., Ramli, A. R., Kahar, B. H. A., & Saripan, M. I. (2009). A frameworkfor white blood cell segmentation in microscopic blood images using digital image processing.Biological Procedures Online, 11(1), 196.

15. Jiang, K., Liao, Q. M., & Dai, S. Y. (2003). A novel white blood cell segmentation schemeusing scale-space filtering and watershed clustering. In Proceedings of the 2003 InternationalConference on Machine Learning and Cybernetics (pp. 2820–2825). IEEE.

16. Torre, V., & Poggio, T. A. (1986). On edge detection. IEEE Transactions on Pattern Analysisand Machine Intelligence, 8(2), 147–163.

17. Acharjya, P. P., Sinha, A., Sarkar, S., Dey, S., & Ghosh, S. (2013). A new approach of watershedalgorithm using distance transform applied to image segmentation. International Journal ofInnovative Research in Computer and Communication Engineering, 1(2), 185–189.

18. Kakarla, J., & Majhi, B. (2013). A new optimal delay and energy efficient coordinationalgorithm for WSAN. In 2013 IEEE International Conference on Advanced Networks andTelecommunications Systems (ANTS) (pp. 1–6).

19. Francos, J. M., Meiri, A. Z., & Porat, B. (1993). A unified texture model based on a 2-D Wold-like decomposition. IEEE Transactions on Signal Processing, 41(8), 2665–2678.

20. Mazalan, S. M. Automated red blood cells counting in peripheral blood smear image usingCircular Hough Transform. In: Artificial intelligence, modelling and simulation (AIMS), firstinternational conference on artificial intelligence, modelling & simulation, IEEE, (2013).

21. Celenk, M. (1990). A color clustering technique for image segmentation. Computer Vision,Graphics, and Image Processing, 52(02), 145–170.

22. Tcheslavski, G. V. (n.d.). Morphological image processing: Basic algorithms. Retrieved fromhttp://ee.lamar.edu/gleb/dip/ (7/05/2016).

http://ee.lamar.edu/gleb/dip/

Smart Assist: Smartphone-Based DrugCompliance for Elderly Peopleand People with Special Needs

Akif Khan and Shah Khusro

1 Introduction

Aging is a natural phenomenon, and the percentage of elderly population andlife expectancy rate of above 65 years are increasing across the globe [1]. Thisincrease in elderly population not only disturbs the balance between earning andotherwise people but also leads to the decline of their healthcare. Aging bringsseveral challenges to elderly people mainly due to chronic age-related diseases likelate learning, acceptability, etc. Increase in the average age of the total populationand the consequent proportional growth in the ascent of diseases have obviousresults on emergency situation in the upcoming years [1].The proportion of personsaged 80 years increased from 7% in 1950 to 14% in 2012. An expected projectionof 19% increase in the population of this age group by 2050 is given.

Elderly persons obviously require support for mobility and extensive care toperform their routine activities. Ambient assisted living (AAL) aims to enhancethe quality of life at their home or living independently in smart houses [2].AAL-oriented solutions are based on ICT technologies to reduce the need forcaretakers and nursing staff. More specifically, AAL is aimed to develop assistancesystems mainly focusing on user-initiated access to services and systems proactivelyprovisioned or in response to the user’s request. These systems sense, capturethe context, and respond intelligently to the minimal interaction of users [3].Ambient intelligence covers various aspects of context-aware computing, disap-pearing computing and pervasive computing for combining features of sensitivity,responsiveness, adaptability, transparency, ubiquity, and intelligence [4]. Thesetechnologies provide an edge to ambient assisted living for senior citizens to achieve

A. Khan · S. Khusro (�)Department of Computer Science, University of Peshawar, Peshawar, Pakistane-mail: [email protected]; [email protected]


99




https://doi.org/10.1007/978-3-319-96139-2_10

100 A. Khan and S. Khusro

comfort, autonomy, and assistance in case of emergency and other scenarios ofchronic nature. On the other hand, smartphones integrate sensing, capturing, storing,and processing capabilities in one portable place and serve as a better alternative tothese traditional devices. These features include high-end processing capabilitiesand integration of diverse nature of sensors, camera, and speech systems and maybe effectively utilized for fall dedication, automated alarm systems, medication, androutine work to social interaction and collaboration.

This article reports on the design and development of augmented reality-basedsmartphone app for drug compliance suited to the needs of elderly people and peoplewith special needs. The contribution will help elderly people to facilitate in takingmedicine on time and in accurate quantity as prescribed by a physician. The aim isto reduce the technicalities to the minimum, which will help not only a senior citizenbut also the least literate person belonging to any age group. The proposed three-layer architecture provides a unified place for drug vendors, hospital administration,people at the pharmacy, doctors, and patients to collectively strive for a betterhealthcare for all. This paper is organized as follows: Section 2 elaborates on thework done so far in this direction, while Sect. 3 describes augmented reality-basedsolutions. Section 3.2 describes the proposed system design and implementation,and Sect. 4 finally concludes the paper with a discussion on future work.

2 Related Work: Toward Smartphone-Based DrugCompliance

In this section, we describe the previously done work on mobile/smartphoneassistance in health sector industry. The smartphone is indeed the success story ofthe decade, and this technology has penetrated significantly in all aspects of society.Healthcare is no exception where it has considerably improved the effectiveness andreduced the cost of traditional medication compliance interventions. Smartphonesare operating different services through dedicated applications called apps. Asmartphone-based medication compliance app can potentially consolidate user-specific drug usage pattern and educate patient for drug usage on specialized timing,dose, and frequency. Many older people also use mobile phones on a regular basisto interact with distant relatives and family and friends for many other purposesincluding networking and socialization [5].

The successful clinical outcome depends on factors like effective monitoring,drug compliance, and personalized medicine management. Samuel et al. [6] haveprovided an analytical review of mobile apps available for Android and iOSplatforms in the area of communication, patient data management, patient healthmonitoring, heart rate assessment, patient compliance, etc. Mobile phone appearsto be the agent of a cultural shift for encouraging old people for seeking newinformation, networking, and bringing autonomy in experiencing the quality oflife [7]. In a short-term intervention, SMS and electronic devices provide the

Smart Assist: Smartphone-Based Drug Compliance for Elderly People. . . 101

patient with an audio or textual reminder as an alert of their medication [8, 9].A number of similar nature studies are being conducted based on smartphoneinterventions in the clinical sessions. However, smartphone-based total adherenceis still an unconquered area [10]. Although mobile phones have now been usedwidely in different contexts and dynamics from the last few decades, smartphonesare a more recent and advanced shape of providing innovative and interactiveservices to the users. Besides the standard facilities of voice, SMS, and phonebook, smartphones also provide advanced features including Internet browsing,running application, geopositioning, etc. Most of the new-generation smartphoneshave excellent usability features like high-resolution display screen, touch screeninteraction, user-friendly user interfaces, etc. The high quality of camera-capturingcapabilities for recording audio/video can facilitate in the utilization of personal careand healthcare applications [11]. Application development is strongly encouragedwith a factor of smartphone high-end computational capabilities, larger memories,large screen display, and extended operating system. Smartphones have achievedpervasive presence in the society in such a way that the ordinary user can find itto self-organize his activities across large geographical area [12]. The enormouspotentials of mobile communication technology can transform healthcare clinicalintervention tremendously.

Previous studies have evaluated the use of mobile phones to support healthcareinterventions in the form of collecting data for healthcare research [13] and wellmolded for utilization of medical and healthcare education for clinical practices.A number of studies have highlighted successful stories of mobile phone usage tosupport remote healthcare in developing nations [14]. Portable handheld computertechnology arises in the area of clinical research. Use of mobile phone records iscompared with traditional paper-based records in controlled drug environments [15].More than 7000 documented cases of smartphone health apps have been reviewedleveraging mobile phone technology for healthcare facilitation [16, 17]. Accordingto a study in the United States, 25% of smartphone users were already using healthapplications, and almost half of those asked would be interested [18]. Smartphoneapps dedicated to drug compliance are extensively reviewed covering all majorsmartphone platforms, i.e., Android, iOS, and BlackBerry. These apps are testedwith the list of desirable attributes where every attribute is given a score dependingon its nature and impact on overall evaluation [19]. These applications were tested intheir corresponding operating systems through a standardized medicine consideringdifferent daily usage medicine of different timing and dose. Apps were furtherevaluated for each manufacturer claim based on the authors’ scoring criteria andfunctionality of the reminder system. Medical reminders were tested for 24 h each.More than 150 applications were tested in the above study.

MyMedSchedule [20] covers compliance for high medication burden includingorgan transplant, HIV, oncology, and hematology. Health professional uses medi-cation databases for the administration of drug intake from the web-based system.These medication intake sequence and dose have pushed to the patient’s mobiledevice notified to the patient via reminder system. Say no to drugs is anothersolution for using a smartphone for taking sufficient amount of drug helping in


drug normalization [21]. Similarly, MedHelp [22] enables demented elderly peopleto improve medication compliance through a Google Glass wearable solution.It proactively engages users through reminders and helps them in recognizingmedicine container. MedLink [23] is a smartphone application that helps in thetreatment of depression cases through a systemic digital intervention to addressand improve quality of primary care. Adherence through this system results in 82%improvement in major depressive disorders. MyMed [24] is a paid service allowinga patient to enter their schedule online. This app is HIPAA-compliant for retrievingand modifying any patient data from the web. Similarly, RxmindMe [25] is alsobased on HIPAA compliance for medication, vitamins, and supplements allowingdosage information and reminder fixing. However, despite these efforts, researchon medication compliance still represents a fundamental healthcare challenge; appsdeveloped for inexpensive, scalable, and accessible to anyone with a smartphoneand do not require separate accessories, which allow the privilege to test easily.

In this paper, we introduce a novel approach called “Smart Assist” whichwas developed to support medication compliance for elderly people. Moreover,we describe its implementation methodology and framework. The system onceimplemented will be tested and evaluated on real data for finding out percentageincrease in the efficacy in drug compliance.

3 Augmented Reality-Based Drug Compliance Application

Mobile augmented reality provides an enriched view of physical world annotatedby digital information. A user can feel, touch, and experience more enrichedpresentation of a physical object, artifacts, and point of interest [26, 27]. It combinesreal and virtual objects in the real environment and registered real and virtual objectsand run interactive in three dimensions, providing freedom for augmenting anyinformation on the live environment. A common scenario of augmented reality hasbeen analyzed for assisting elderly people to perform a number of tasks on a dailybasis.

3.1 Scenario: Drug Intake via Object Recognition/ComputerVision

Four operations are required for achieving augmented reality experience usingimage capturing and object recognition techniques [27, 28]. Scene capturing is thefirst point where the camera captures a picture of an object or place of interest,which is stored locally and compared with or scanned in real time with the availabledata store of images. The scene is identified based on prior registration of the objectin the augmented reality system, while the scene is processed based on augmented


Fig. 1 Augmentation process

Fig. 2 Annotation process

reality scenario and resulting visual object representation, or the visual presentationis provided to the end user on top of the camera scene. Other than the direct objectidentification and processing, the visual tag can play a vital role in simplifying theprocess. A QR code [29] is machine-readable optical label containing informationabout the item to which it is associated. QR codes are gaining popularity inindustries like automobile and marketing, due to its features of fast readability andstorage capacity. Similarly, it has vital spectrum open for usage in the domain ofambient assisted living. Medication, home labeling, easy shopping, and identifyingcalories in food products are a few examples used in ambient assisted living domain.Hervas et al. [30] demonstrated an augmented reality-based medication case studyillustrated in Figs. 1 and 2, respectively. Figure 1 outlooks object scan-to-augmentprocess, while Fig. 2 illustrates the final output of augmented solution on top ofthe camera view as rich annotation contents. This case study demonstrates a usableexperience for taking medicine on prescribed time with appropriate frequency.


3.2 System Design and Implementation

Smart Assist aims at building semantically enriched augmented reality-based drugcompliance solution for elderly people in recognizing medication container, takingappropriate dose frequency, and on-time medicine intake. The scope of the solutionis initially confined to primary patient care, clinician involvement, and pharmacyvendors. The main functionality of Smart Assist platform is to provide a seamless,user-friendly, and highly interactive application through smartphone camera.

The elderly people may use the application for drug administration, medicineintake, dose frequency, and reminders for compliance. Built-in camera of the smart-phone is used to recognize the medicine container and medicines; subsequentlythe application will guide the users on taking appropriate dose and frequencyon medicine intake. The application will help in informing the users about theinformation about the remaining drugs left over and remind him/her accordingly.This prototype is developed for Android platform. The application has been writtenin Java language for Android Dalvik virtual machine. OpenCV library is usedfor scanning/recognizing drug images and comparing with original stored in thedatabase. SQLite database is used to store in-device information about drugs,captured images, etc., whereas, through dedicated web services, the entire system isconnected for gaining a unified experience and clinical analytics. The interface withthe clinicians’ site has been accomplished using the W3C Web Services technology,while the interface with the patient is mainly through the Android app. Variouscomponents of this architecture are illustrated below.

Generic Architecture of Smart Assist

Smart Assist is a generalized three-layer framework for augmented assistance toelderly people in achieving the goal of drug compliance (shown in Fig. 3) consistingof user interface layer, cloud entry points, clinicians’ sites, and pharmacy vendormanagement. Respective components are represented in the form of specializeduser interface integrated with cloud endpoints. The integration and information flowamong various components of the framework are depicted in Fig. 3.

Pharmacy Vendor Layer

Pharmacy vendor management is a web-/mobile-based interface connected withcentral web cloud of drug compliance app. This component provides drug details,chemical composition, precautions, etc. of particular drugs available with therespected pharmaceutical vendor. The primary role of this component is uploadingimages of drug packaging, tablet packing, etc. from different angles. These imageswill be utilized by patient drug compliance mobile app and through computer visionapproach and recognition subsystem for annotation purpose.


Fig. 3 Smart Assist architecture

Clinical Administration Layer

Clinical administration and the prescription layer are web/mobile interface and willbe connected with central cloud of drug compliance app. Hospital medical staffwill be having access to this component. Concerned medical physician will set upa prescription and trigger a sequence of medication, drug intake time, and remindersetting (date/time) for a patient registered through MR no (medical record number,unique identification of patient). This component can also be used in the future foradvanced analytics of determining patient compliance and noncompliance ratio andthe degree of improvement for taking particular dose. The doctor can also adjustmedicine in the next visit of the patient as required, as every information will berecorded in drug compliance cloud so that information can be consolidated on everyrevisit of the patient to the hospital.

Patient Drug Compliance Layer

Patient drug compliance layer is the most important part of the overall solution. Thepurpose solution is to provide the older-aged people to take their medicine on timeand within prescribed dose limit. The patient smartphone will act as smart assistantto help in taking drugs and ensuring compliance of drugs. Patient smartphone willconfigure to his/her MR record connected via web services of drug compliancecloud. A set of prescription tags and models can be downloaded on a patientsmartphone, for instance set of images of drugs being used in his/her prescription


Fig. 4 Scan drug flow diagram and Smart Assist app

and dose frequency and sequence of drug intake. The patient can scan drug as shownin Fig. 4 using computer vision technique of augmented reality technology. Thesystem provides an annotated, enriched presentation of administrating medicine.The smartphone after recognizing the image of the drug will update the drug charttable, and further frequency will be automatically adjusted as per plan in clinicalprescription.

4 Conclusion

Augmented reality-based solutions on Android-based devices are believed to havea significant impact in experiencing improved quality of life for elderly people.Smartphone-based interventions have a great impact on drug compliance andto enhance the productivity in medication management. Due to the success ofthese interventions, a widespread acceptance is reported from the society as wellas from the patient perspective. This has also increased the public expectationtoward a positive attitude of adopting mobile-based healthcare services. Chronicdiseases can be controlled efficiently, if monitored and complied with proper dugcompliance systems and applications. Presently, health-related mobile computingapplications are playing its role for elderly and chronically ill users mainly due tointeractive presentation usable user interfaces and accessibility features. Most ofthe commercially successful applications are likely to be targeting younger people(between 20 and 40 years) and healthy individuals. However, the young and middle-aged of today is the senior citizen of the tomorrow. This effect will alter the entirespectrum of ICT-based services for performing their daily life tasks. Smart Assisthas expected to fulfill the unrestricted conditions of healthcare applications, and


this was the motive behind this contribution to make a healthy society for elderlyaged people. This paper is focused on an augmented reality-based solution fordrug compliance for elderly aged people; with this, now the patients have greatercontrol over medication management in natural and flexible way. Major endpointsof the system have been enriched by their respective web/mobile interfaces; thesystem is connected with central drug compliance cloud. Patient viewpoint, clinicalviewpoint, and pharmacy vendor management are extensive subsystems performingas per drug compliance business logic. Data store of central cloud and device-specific data are stored separately. However, synchronization of data takes placeafter a set of intervals. A complete history of patient records and prescription ismaintained for future analytics in drug compliance and medication management.The proposed system was tested for few cases, which resulted in considerably bettermedication compliance ratio. However, the in-depth analysis may be required to teston large-scale setups by integrating hospitals, pharmacy vendors, and smart homesfor close and effective results.

In terms of future work, we aim to outline an analytics for drug compliance to beimplemented in the smart home, nursery care for elderly people. Our future workwill be more focused on unified, easy-to-operate user interfaces for elderly peopleand blind people. The same application will be tailored closely to specific demandsof elderly people and blind people to accept as part of their complicated day-to-daylife routine. An emergency response notification system will be added to this SmartAssist app for handling over-dose and under-dose management in case of an acuteemergency.

References

1. UNO, World population ageing 2013. Available from http://www.un.org/en/development/desa/population/publications/pdf/ageing/WorldPopulationAgeing2013.pdf

2. O’Mullane, B., Knapp, R. B., & Bond, R. (2017). Review of user interface devices for ambientassisted living smart homes for older people. Gerontechnology, 9(2), 319.

3. Nehmer, J., Becker, M., Karshmer, A., & Lamm, R. Living assistance systems: An ambientintelligence approach. In Proceedings of the 28th international conference on Softwareengineering, ACM, pp. 43–50.

4. Rodrigues, G. N., Alves, V., Silveira, R., & Laranjeira, L. A. (2012). Dependability analysisin the ambient assisted living domain: An exploratory case study. Journal of Systems andSoftware, 85(1), 112–131.

5. Katz, J. E., Katz, J. E., & Aakhus, M. (2002). Perpetual contact: Mobile communication,private talk, public performance. Cambridge, UK: Cambridge University Press.

6. Gan, S., Koshy, C., Nguyen, P. V., & Haw, Y. X. (2016). An overview of clinically andhealthcare related apps in Google and Apple app stores: Connecting patients, drugs, andclinicians. Scientific Phone Apps and Mobile Devices, 2(1), 1–9.

7. Jenkins, H. (2006). Convergence culture: Where old and new media collide. New York: NYUpress.

8. Vervloet, M., Linn, A. J., van Weert, J. C., De Bakker, D. H., Bouvy, M. L., & Van Dijk, L.(2012). The effectiveness of interventions using electronic reminders to improve adherence tochronic medication: A systematic review of the literature. Journal of the American MedicalInformatics Association, 19(5), 696–704.

http://www.un.org/en/development/desa/population/publications/pdf/ageing/WorldPopulationAgeing2013.pdf


9. Noh, J.-H., Cho, Y.-J., Nam, H.-W., Kim, J.-H., Kim, D.-J., Yoo, H.-S., Kwon, Y.-W., Woo,M.-H., Cho, J.-W., & Hong, M.-H. (2010). Web-based comprehensive information system forself-management of diabetes mellitus. Diabetes Technology & Therapeutics, 12, 333–337.

10. Wu, R., Rossos, P., Quan, S., Reeves, S., Lo, V., Wong, B., Cheung, M., & Morra, D. (2011).An evaluation of the use of smartphones to communicate between clinicians: A mixed-methodsstudy. Journal of Medical Internet Research, 13, e59.

11. Ranganathan, L. N., et al. (2015). Application of mobile phones in epilepsy care. InternationalJournal of Epilepsy, 2(1), 28–37.

12. Rheingold, H. (2007). Smart mobs: The next social revolution. USA: Basic books, e-Book.13. Blaya, J. A., Fraser, H. S., & Holt, B. (2010). E-health technologies show promise in developing

countries. Health Affairs, 29, 244–251.14. Kaplan, W. A. (2006). Can the ubiquitous power of mobile phones be used to improve health

outcomes in developing countries. Globalization and Health, 2, 1–14.15. Lane, S. J., Heddle, N. M., Arnold, E., & Walker, I. (2006). A review of randomized controlled

trials comparing the effectiveness of hand held computers with paper methods for datacollection. BMC Medical Informatics and Decision Making, 6, 23.

16. Phillips, G., Felix, L., Galli, L., Patel, V., & Edwards, P. (2010). The effectiveness of m-healthtechnologies for improving health and health services: A systematic review protocol. BMCResearch Notes, 3, 250.

17. Terry, M. (2010). Medical apps for smartphones. Telemedicine Journal and e-Health, 16, 17–22.

18. Fronstin, P. (2011). Findings from the 2011 EBRI/MGA consumer engagement in health caresurvey. EBRI Issue Brief, (365), 1–26.

19. Dayer, L., Heldenbrand, S., Anderson, P., Gubbins, P. O., & Martin, B. C. (2013). Smartphonemedication adherence apps: Potential benefits to patients and providers. Journal of theAmerican Pharmacists Association: JAPhA, 53, 172.

20. MyMedSchedule. (2014). Available from: http://www.medactionplan.com21. Zhang, M. W., & Ho, R. C. (2016). Smartphone for the smarter delivery of drugs, psychoeduca-

tional materials and acute intervention for at-risk users. BMJ Innovations, 2016(2), 136–138.22. Xu, Q., et al. (2016). MedHelp: Enhancing medication compliance for demented elderly people

with wearable visual intelligence. Scientific Phone Apps and Mobile Devices, 2(1), 1–4.23. Corden, M. E., et al. (2016). MedLink: A mobile intervention to improve medication adherence

and processes of care for treatment of depression in general medicine. Digital Health, 2,2055207616663069.

24. MyMed. (2014). Available from: http://www.my-meds.com25. Rxmindme. (2014). Available from: https://itunes.apple.com/pk/app/rxmindme-prescription-

medicine26. Azuma, R., Baillot, Y., Behringer, R., Feiner, S., Julier, S., & MacIntyre, B. (2001). Recent

advances in augmented reality. Computer Graphics and Applications, IEEE, 21, 34–47.27. Khan, A., et al. (2015). Rebirth of augmented reality-enhancing reality via smartphones. Bahria

University Journal of Information & Communication Technology, 8(1), 110.28. López, H., Navarro, A., & Relaño, J. An analysis of augmented reality systems. In Computing

in the Global Information Technology (ICCGI), 2010 Fifth International Multi-Conference on,IEEE, pp. 245–250.

29. Falas, T., & Kashani, H. Two-dimensional bar-code decoding with camera-equipped mobilephones. In Pervasive Computing and Communications Workshops, 2007 PerCom Workshops’07 Fifth Annual IEEE International Conference on, IEEE, pp. 597–600.

30. Hervás, R., Garcia-Lillo, A., & Bravo, J. (2011). Mobile augmented reality based on thesemantic web applied to ambient assisted living. In International workshop on ambient assistedliving (pp. 17–24). Berlin, Heidelberg: Springer.

http://www.medactionplan.com/

http://www.my-meds.com

https://itunes.apple.com/pk/app/rxmindme-prescription-medicine

An Overview of OCT Techniques forDetection of Ophthalmic Syndromes

Adeel M. Syed, Muhammad Usman Akbar, and Joddat Fatima

1 Introduction

Optical coherence tomography (OCT) is a recognized technique for medical imag-ing that makes use of light to capture very high resolution [1]. The images acquiredthrough OCT are in three-dimensional, and these images are from within opticalscattering media [2]. These images provide cross-sectional view of the eye. Due toits robustness and efficiency, it has become an integral part of the examination of theeye. OCT has more than one type, and with the introduction of Fourier domain, anewer type of OCT imaging facility, the speed and accuracy of the OCT images, hasbeen increased along with the resolution of the image gathered [3]. More work isbeing done, and continuous improvement is under process in the quality and speedof the OCT image acquisition. Researchers are trying to increase the line rate [4]to a maximum possible. They are also trying their best to reduce the overall scantime [5, 6]. From the whole process, once they have acquired the high-resolutionimages, they want to perform different morphological measurements [7] so that anunderstanding can be derived.

OCT is able to gather images in segments and slices. The front-most sliceor segment, also termed as anterior segment, and the rear segment, also knownas posterior segment images, can be acquired using an OCT imaging machine,and analysis can be performed upon them [3]. The diseases, diabetic retinopathy,pathological myopia, and glaucoma, can be detected using these OCT images [8,9]. By the passage of time, new progressions in OCT image apprehending apparatusare being done, and now, the devices are strong enough to go to the deeper structuresof the eye such as till the optical nerve. Patients with glaucoma have a lower

A. M. Syed · M. U. Akbar · J. Fatima (�)Bahria University Islamabad, Islamabad, Pakistan


109


https://doi.org/10.1007/978-3-319-96139-2_11

110 A. M. Syed et al.

average retinal nerve fiber layer (RNFL) [10], and it can work as a biomarker, witha characteristic by which a specific pathological or physiological disease can berecognized. Therefore, it can be said that the biggest application of OCT is detectionof macular and retinal disorders [11].

Because of the sensitivity and robustness offered by OCT technique, it ispredicted that OCT will be used as the primary mechanism for detection andthroughout examination of a patient of any eye disease patient. OCT has the abilityto detect the biology of the retina, and the work also explains the production anddevelopment of the retinal damages that are the leading causes of blindness andnear to complete vision loss. It is predicted that OCT will prove to be an importantsource of collection of data and perform investigation of any sort of disease ofthe eye; however it shall not completely replace the existing ophthalmic imagingtechniques. Along with that, the OCT will significantly increase the research in thefield, and different quantitative and qualitative data would be able to be gatheredusing some morphological operations that can be performed in the images gatheredby OCT [2].

The paper is organized in such a way that in the next section, Sect. 2, weare going to explain the working principles and some of the basic terminologies.In Sect. 3, we reported the causes and symptoms of disease, and performed thepathologies identification in an OCT image. In Sect. 4, we analyzed differentcomputer vision and image processing techniques for identification of the diseasesand retinal disorders. Finally, conclusion and future research directions are given inthe last section.

2 OCT Terminologies and Modalities

2.1 OCT Terminologies

OCT works with light-like ultrasound works with sound. It measures the delay ofreflected radiations, and that’s how it works. Crosswise direction and longitudinaldepth and some scanning modalities are used to characterize the cross-sectionaldata. These scanning modalities have been classified in three categories. The namesof these categories are kept similar to the ones in the ultrasound technology. Thenames are A scan, B scan, and C scan [12]. A scan in the ultrasound, also knownas axial scan, records two important factors of the sound echo received. These arethe amplitude and the time the pulse takes to complete one complete round trip. Oncontrary, in the OCT systems, an A scan can be measure by scanning the whole eyein longitudinal manner. The process is shown in Fig. 1. Figure 2 shows how A scanresults in an image frame acquisition. The B scans uses A scans, several in number,in longitudinal and lateral scanning. The collection of several B scans can be used tocreate a three-dimensional dataset [13]. Lastly, the C scancan be acquired by a step

An Overview of OCT Techniques for Detection of Ophthalmic Syndromes 111

Fig. 1 Scanning of whole eye in longitudinal way

Fig. 2 A scan, B scan, and 3D scan

further in B scan. In B scanning, the lateral scanning is done in a different manner,and rather two-dimensional lateral scanning are done by several successive scansin x and y directions. This provides a cross-sectional image of the compete eye atdifferent depths.


2.2 OCT Modalities

The OCT has two ways to function. It can either work in time domain (TD OCT) orFourier domain (FD OCT) principle [14]. These two are further explained in detailin the subsequent sections.

2.3 Time Domain OCT Systems (TD OCT)

In time domain OCT systems, the sample is placed on one end, and a light sourcethrows light upon it from the other end. There is a place between the light sourceand the object where a beam splitter (BS) is placed. The beam splitter divides thelight in two segments. A part of light is allowed to travel straight to the object,while a part of it is deflected toward a reference mirror. The light from both ends getreflected back and reaches the beam splitter. Both the components are remerged andare forwarded toward and detected by a photo detector. This is how a simple timedomain OCT imaging hardware works in principle.

When OCT was invented, the time domain OCT was the first used methodology.Later on, as time passed, a newer method was discovered that was named as Fourierdomain OCT. The key difference between the two methods is in TD OCT; themoving mirror at the reference arm results in slower image acquisition rate [13].It is explained in detail in the following subsection. Comparing the speed, the TDOCT can capture approximately 450 A scans per second [15], whereas the FD OCTcan capture approximately 55,000 A scans per second [16] proving it to be morethan 100 times faster. Image quality of TD OCT is also not so good, and it is shownin Fig. 3.

Fig. 3 Time domain OCTimage


Fig. 4 Different categoriesof OCT OCT

TD OCT FD OCT

SD OCT SS OCT

2.4 Fourier Domain OCT Systems (FD OCT)

Fourier domain OCT is further subdivided into two types. These are spectral domainOCT (SD OCT) and swept source OCT (SS OCT). The basic tree structure of ourclassification of types of OCT is shown in Fig. 4. The lateral subdivision of FD OCTboth allow three-dimensional imaging of the frame. For example, as explained in theabove text of the last subsection, the scanning is done in longitudinal as well as in2D lateral scanning combined to make it a three-dimensional scan overall. SD OCTand SS OCT are further explained in the subsections below.

2.5 Spectral Domain OCT (SD OCT) SD OCT

From [17] we have concluded that a spectral domain OCT provides better resolutionand a much faster scan time. Up to 15,000 A scans per second can be achieved [18].Later on, this speed increased to 18,000 A scans per second [19] with some minorhardware tweaks. Next year, the number increased to 30,000 A scans per second[20], and within the next 3 years, it reached to 50,000 A scans per second [21]. Themajor breakthrough in the scanning speed came in 2008 by a research group of MITwho used a high-speed scan camera with a CMOS detector achieving a speed ofmore than 300,000 A scans per second [22]. In order to keep the data in order withthe other sources available, range more than 1 mm and a resolution of approximately2 μm and 120,000 scans per second [23].

2.6 Swept Source OCT (SS OCT)

SS OCT makes use of an extraordinary bandwidth PD and a frequency swept laser[24]. OFDI, optical Fourier domain imaging, refers to the same thing [25]. In [26,27] different technical details are discussed telling the technicalities about sweptsource OCT. 15 kHz–115 kHz frequency rages are used for gathering the OCToutputs [28]. Commercially it is available at a frequency rate of 40 kHz [29]. Talking


Table 1 Comparison table

Name Resolution A scans/sMax num for Bscan Additional feature

Nidek RS-3000 7 μm20 μm 53,000 1024 SLOOptoveu OCTscanner

5 μm15 μm 26,000 1024 SLO, color

Zeiss cirrus 4000 5 μm15 μm 27,000 4096 SLOTopcon 2000 5-6 μm<20 μm 27,000 4096 Infrared charge-coupled

deviceOpko/OTI 6 μm15 μm 27,000 1024 SLOHeidelbergSpectralis

7 μm14 μm 40,000 4096 SLO, red free, andfluorescein angiography

about the speed of the SS OCT outputs, 50,000 A scans per second can be acquired[30]. In [31] researches have shown outputs with different rates and frequencies ofthe OCT devices available in the market.

The table shows the information gathered from and shows different devicesavailable internationally that provide the services of FD OCT image acquisition.All machines are able to capture OCT and fundus images; however some of themachines are also able to capture SLO images (Table 1).


OCT images can be used in TD OCT or FD OCT techniques. FD OCT is able togenerate more crisp and high-resolution outputs with lesser processing time. FDOCT is able to generate true 3D figures; however more research can be done in thisfield [9]. To date, for the detection of glaucoma, SD OCT images are used moreoften; however our research indicates that more work can be done in the same usingSS OCT images. In conclusion, we can indicate that OCT images can serve as avery informative source of detection of diseases being caused by glaucoma, macularedema, and the like. There is a huge research gap in using the SS OCT images forcomputer-aided algorithmic detection, and more work will be done in this field inthe future.

References

1. Huang, D., Swanson, E. A., Lin, C. P., Schuman, J. S., Stinso, W. G., Chang, W., Hee, M.R., Flotte, T., Gregory, K., Puliafito, C. A., & Fujimoto, J. G. (1991). Optical coherencetomography. Science, 254(5035), 1178–1181.

2. Drexler, W., & Fujimoto, J. G. (2008). State-of-the-art retinal opticalcoherence tomography. Progress in Retinal and Eye Research, 27(1), 45–88.https://doi.org/10.1016/j.preteyeres.2007.07.005.

http://dx.doi.org/10.1016/j.preteyeres.2007.07.005


3. Poddar, R., & Reddikumar, M. (2015). In vitro 3D anterior segment imaging in lamb eyewith extended depth range swept source optical coherence tomography. Optical and LaserTechnology, 67, 33–37. https://doi.org/10.1016/j.optlastec.2014.09.007.

4. Povazay, B., Hofer, B., Torti, C., Hermann, B., Tumlinson, A. R., Esmaeelpour, M., Egan, C.A., Bird, A. C., & Drexler, W. (2009). Impact of enhanced resolution, speed and penetrationon three-dimensional retinal optical coherence tomography. Optics Express, 17(5), 4134–4150.https://doi.org/10.1364/OE.17.004134.

5. Lebed, E., Mackenzie, P. J., Sarunic, M. V., & Beg, F. M. (2010). Rapid volumetricOCT image acquisition using compressive sampling. Optics Express, 18(20), 21003–21012.https://doi.org/10.1364/OE.18.021003.

6. Young, M., Lebed, E., Jian, Y., Mackenzie, P. J., Beg, M. F., & Sarunic, M. V. (2011). Real-timehigh-speed volumetric imaging using compressive sampling optical coherence tomography.Biomedical Optics Express, 2(9), 2690–2697. https://doi.org/10.1364/BOE.2.002690.

7. Sieun, L., Lebed, E., Sarunic, M. V., & Beg, M. F. (2015). Exact surface registration of retinalsurfaces from 3-D optical coherence tomography images. IEEE Transactions on BiomedicalEngineering, 62(2), 609–617. https://doi.org/10.1109/TBME.2014.2361778.

8. Puliafito, C. A., Hee, M. R., Lin, C. P., Reichel, E., Schuman, J. S., Duker, J. S., Izatt, J. A.,Swanson, E. A., & Fujimoto, J. G. (1995). Imaging of macular diseases with optical coherencetomography. Ophthalmology, 102(2), 217–229.

9. Abramoff, M. D., Garvin, M. K., & Sonka, M. (2010). Retinal imagingand image analysis. IEEE Reviews in Biomedical Engineering, 3, 169–208.https://doi.org/10.1109/RBME.2010.2084567.

10. Bron, A. M., Francoz, A., Beynat, J., Nicot, F., Cattaneo, A., & Creuzot, C. (2011). Is choroidalthickness different between glaucoma patients and healthy subjects? Acta Ophthalmologica.https://doi.org/10.1111/j.1755-3768.2011.4353.x.

11. Bowd, C., Zangwill, L. M., Blumenthal, E. Z., Vasile, C., Boehm, A. G., Gokhale, P. A.,Mohammadi, K., Amini, P., Sankary, T. M., & Weinreb, R. N. (2002). Imaging of the opticdisc and retinal nerve fiber layer: The effects of age, optic disc area, refractive error, andgender. Journal of the Optical Society of America. A, Optics, Image Science, and Vision, 19(1),197–207.

12. Bouma, B. (2001). Handbook of optical coherence tomography. New York: Taylor & Francis.13. Schuman, J. S. (2008). Spectral domain optical coherence tomography for glaucoma (an AOS

thesis). Transactions of the American Ophthalmological Society, 106, 426–458.14. Wang, R. K., & Tuchin, V. V. (2013). Advanced biophotonics: Tissue optical sectioning. New

York: Taylor & Francis.15. Hee, M. R., Izatt, J. A., Swanson, E. A., Huang, D., Schuman, J. S., Lin, C. P., Puliafito, C.

A., & Fujimoto, J. G. (1995). Optical coherence tomography of the human retina. Archives ofOphthalmology, 113(3), 325–332.

16. Wojtkowski, M., Srinivasan, V., Fujimoto, J. G., Ko, T., Schuman, J. S., Kowal-czyk, A., & Duker, J. S. (2005). Three-dimensional retinal imaging with high-speedultrahigh-resolution optical coherence tomography. Ophthalmology, 112(10), 1734–1746.https://doi.org/10.1016/j.ophtha.2005.05.023.

17. Wojtkowski, M. (2010). High-speed optical coherence tomography: Basics and applications.Applied Optics, 49(16), D30–D61. https://doi.org/10.1364/AO.49.000D30.

18. Wojtkowski, M., Bajraszewski, T., Targowski, P., & Kowalczyk, A. (2003). Real-time invivo imaging by high-speed spectral optical coherence tomography. Optics Letters, 28(19),1745–1747. https://doi.org/10.1364/OL.28.001745.

19. Yun, S., Tearney, G., Bouma, B., Park, B., & de Boer, J. (2003). Highspeed spectral-domainoptical coherence tomography at 1.3 lm wavelength. Optics Express, 11(26), 3598–3604.https://doi.org/10.1364/OE.11.003598.

20. Nassif, N., Cense, B., Hyle Park, B., Yun, S. H., Chen, T. C., Bouma, B. E., Tear-ney, G. J., & de Boer, J. F. (2004). In vivo human retinal imaging by ultrahigh-speed spectral domain optical coherence tomography. Optics Letters, 29(5), 480–482.https://doi.org/10.1364/OL.29.000480.

http://dx.doi.org/10.1016/j.optlastec.2014.09.007

http://dx.doi.org/10.1364/OE.17.004134

http://dx.doi.org/10.1364/OE.18.021003

http://dx.doi.org/10.1364/BOE.2.002690

http://dx.doi.org/10.1109/TBME.2014.2361778

http://dx.doi.org/10.1109/RBME.2010.2084567

http://dx.doi.org/10.1111/j.1755-3768.2011.4353.x

http://dx.doi.org/10.1016/j.ophtha.2005.05.023

http://dx.doi.org/10.1364/AO.49.000D30

http://dx.doi.org/10.1364/OL.28.001745

http://dx.doi.org/10.1364/OE.11.003598

http://dx.doi.org/10.1364/OL.29.000480


21. Bajraszewski, T., Wojtkowski, M., Szkulmowski, M., Szkulmowska, A., Huber, R., & Kowal-czyk, A. (2008). Improved spectral optical coherence tomography using optical frequencycomb. Optics Express, 16(6), 4163–4176. https://doi.org/10.1364/OE.16.004163.

22. Potsaid, B., Gorczynska, I., Srinivasan, V. J., Chen, Y., Jiang, J., Cable, A., & Fuji-moto, J. G. (2008). Ultrahigh speed spectral/Fourier domain OCT ophthalmic imagingat 70,000 to 312,500 axial scans per second. Optics Express, 16(19), 15149–15169.https://doi.org/10.1364/OE.16.015149.

23. Grulkowski, I., Gora, M., Szkulmowski, M., Gorczynska, I., Szlag, D., Marcos, S.,Kowalczyk, A., & Wojtkowski, M. (2009). Anterior segment imaging with spectralOCT system using a high-speed CMOS camera. Optics Express, 17(6), 4842–4858.https://doi.org/10.1364/OE.17.004842.

24. Choma, M., Sarunic, M., Yang, C., & Izatt, J. (2003). Sensitivity advantage of swept sourceand Fourier domain optical coherence tomography. Optics Express, 11(18), 2183–2189.https://doi.org/10.1364/OE.11.002183.

25. Yun, S. H., Boudoux, C., Tearney, G. J., & Bouma, B. E. (2003). High speed wavelength-sweptsemiconductor laser with a polygon scanner based wavelength filter. Optics Letters, 28(20),1981–1983. https://doi.org/10.1364/OL.28.001981.

26. Golubovic, B., Bouma, B. E., Tearney, G. J., & Fujimoto, J. G. (1997). Optical frequency-domain reflectometry using rapid wavelength tuning of a Cr4? : Forsterite laser. Optics Letters,2(22), 1704–1706. https://doi.org/10.1364/OL.22.001704.

27. Yun, S., Tearney, G., de Boer, J., Iftimia, N., & Bouma, B. (2003). Highspeed optical frequency-domain imaging. Optics Express, 11(22), 2953–2963.https://doi.org/10.1364/OE.11.002953.

28. Oh, W. Y., Yun, S. H., Tearney, G. J., & Bouma, B. E. (2005). 115 kHz tuning repetition rateultrahigh-speed wavelength-swept semiconductor laser. Optics Letters, 30(23), 3159–3161.https://doi.org/10.1364/OL.30.003159.

29. Huber, R., Wojtkowski, M., Fujimoto, J. G., Jiang, J. Y., & Cable, A. E. (2005). Three-dimensional and C-mode OCT imaging with a compact, frequency swept laser source at 1300nm. Optics Express, 13(26), 10523–10538. https://doi.org/10.1364/OPEX.13.010523.

30. Larina, I. V., Furushima, K., Dickinson, M. E., Behringer, R. R., & Larin, K. V. (2009). Liveimaging of rat embryos with Doppler swept source optical coherence tomography. Journal ofBiomedical Optics, 14(5), 050506–050503. https://doi.org/10.1117/1.3241044.

31. Choma, M. A., Hsu, K., & Izatt, J. A. (2005). Swept source optical coherence tomographyusing an all-fiber 1300-nm ring laser source. Journal of Biomedical Optics, 10(4), 044009–044006. https://doi.org/10.1117/1.1961474.

http://dx.doi.org/10.1364/OE.16.004163

http://dx.doi.org/10.1364/OE.16.015149

http://dx.doi.org/10.1364/OE.17.004842

http://dx.doi.org/10.1364/OE.11.002183

http://dx.doi.org/10.1364/OL.28.001981

http://dx.doi.org/10.1364/OL.22.001704

http://dx.doi.org/10.1364/OE.11.002953

http://dx.doi.org/10.1364/OL.30.003159

http://dx.doi.org/10.1364/OPEX.13.010523

http://dx.doi.org/10.1117/1.3241044

http://dx.doi.org/10.1117/1.1961474

Fully Automated Identification of HeartSounds for the Analysisof Cardiovascular Pathology

Ghafoor Sidra, Nasim Ammara, Hassan Taimur, Hassan Bilal,and Ahmed Ramsha

1 Introduction

Nowadays, heart attacks are one of the major causes of mortality in a society.As per WHO, Pakistan is ranked at 63rd in having CVD [1]. Apart from this,17.5 million people die each year due to CVD, which covers 31% of total deathsacross the globe [1]. In the USA, about 610,000 people die each year due to heartdiseases [2], and heart diseases affect approximately 9.3% of the US population[3]. So, in this perilous situation, early detection of CVD would be very beneficialto save as many lives as possible. There are three commonly used techniques todetect CVD. First one is ECG which is the oldest and efficient technique for thedetection of heart rates. However in remote areas, establishing ECG machines is notconceivable [4]. The other technique is photoplethysmogram (PPG) which entailsa long setting time. The third one is PCG; it epitomizes the complete informationabout the heartbeats. PCG signals are highly cost-effective, and they are acquiredthrough digital stethoscopes. Cardiologists then give a verdict on acquired data withhis or her experience. But in remote areas, cardiologists are rarely available, andtherefore people living in rural areas are unable to observe their heart conditionsmore frequently. PCG signal is generally produced during each cardiac cycle whenblood flows through heart valves. It produces four sounds, namely, S1, S2, S3, and

G. Sidra · N. AmmaraDepartment of Electrical Engineering, Bahria University, Islamabad, Pakistan

H. TaimurDepartment of GeoGraphix R&D, LMKR (Pvt.) Limited, Islamabad, Pakistane-mail: [email protected]

H. Bilal · A. Ramsha (�)Department of Electrical Engineering, National University of Sciences and Technology,Islamabad, Pakistan


117



https://doi.org/10.1007/978-3-319-96139-2_12

118 G. Sidra et al.

Fig. 1 PCG signals of (a) healthy subject and (b) diseased subject

S4 [5]. S1 and S2 are the common heart sounds which are also known as lub-dub.The PCG signal of healthy subject and cardiac patient is shown in Fig. 1.

Many researchers have done clinical work related to auscultation of PCG signals,and some of them have also worked on automatically identifying CVD from PCGsignals. K. Agrawal et al. [6] proposed wavelet transform-based technique thatincorporates optimum subband thresholding procedure, and this procedure is used inremoving the noise from highly corrupted signals. N. Shankar et al. [7] used signalprocessing module which contains four blocks, i.e., data acquisition, segmentation,feature extraction, and murmur detection, and by using all these steps, it revealsimportant information about CVD disorders. Heart sound analyzer consists of dataacquisition at different positions, analysis of signal by extracting information, andprograms based on acknowledgment to deliver a likely diagnosis [8]. D. S. V. Sankaret al. [9] proposed a method which is based on principal component analysis (PCA)for segmentation of heart sounds in which they extracted envelop feature set byusing Shannon energy and segmented the candidate signal through cardiac epochs.Some researchers extracted features with a novel tactic in which signal temporaldynamics are modeled using Markov chain analysis and corrected segmentationof heart sounds with real-world noisy PCG signals through hidden semi-Markovmodel (HSMM) [10, 11]. J. Pedrosa et al. [12] proposed two novel algorithms.The first one is related to the segmentation of the heartbeats, and it is based onautocorrelation. The second one is related to the detection of heart murmurs, and itis based on the collection of features from miscellaneous domains. P. Langley et al.[13] computed wavelet entropy to test the classification feasibility of unsegmentedand short-duration recordings. K. Suhas et al. [14] used SVM classifier and achievedthe accuracy of around 72.70%. S. Yazdani et al. [15] classified murmur soundsbased on beat-extracted features and tape-long feature analysis from PCG signals.Their achieved sensitivity with 82 samples was 89%. D. B. Springer et al. [16]

Fully Automated Identification of Heart Sounds for the Analysis. . . 119

developed an automated method to measure the quality of mobile phone-recordedheart sound signals using logistic regression classification model. They were able toclassify good- and bad-quality mobile phone-recorded and electronic stethoscope-recorded signals with the accuracy of 82.2% and 86.5%, respectively. Here, weproposed a fully automated system for the identification of heart sounds that areused to automatically diagnose the cardiac abnormalities. The proposed system isrobust, and it takes around 3 s on average to predict the cardiac pathology on amachine with core i5 processor and 4GB DDR3 RAM. The proposed system wastested on a publicly available PASCAL dataset from which it correctly classified 48out of 55 cardiac pathologies with the accuracy, sensitivity, specificity, and negativepredictive value (NPV) of 87.2%, 96.7%, 75%, and 94.7%, respectively.


This paper proposes a fully automated heart sound identification system for theclassification of cardiac pathology. The proposed system is based on three stages.In the first stage, the PCG signal is loaded into the proposed system where it isde-noised through Savitzky-Golay filter. Afterward, lub-dub sounds are extractedfrom the de-noised signal by decomposing it into multiple spectral componentsthrough Daubechies 2-based wavelet decomposition. The extracted lub-dub soundsare passed to the second stage of the proposed system where five distinct featuresare computed which are passed to the third stage of the proposed system to predictthe underlying cardiovascular pathology through an ensemble of supervised SVM,KNN, and Naïve Bayes classifiers. Figure 2 shows the detailed block diagram of theproposed system.

2.1 PCG Dataset Description

We have used publicly available PASCAL dataset in this research from which weconsidered 55 PCG signals of 24 healthy subjects and 31 diseased subjects. Thedataset has been annotated by multiple expert cardiologists, and we have comparedthe performance of the proposed system with those annotations. The dataset can beaccessed through the following link: http://www.peterjbentley.com/heartchallenge/[17].

2.2 Preprocessing

When the candidate PCG signal is loaded into the proposed system, it is de-noised.The reason for de-noising the candidate signal is to remove unwanted artifacts which


120 G. Sidra et al.

Fig. 2 Block diagram of proposed system

were collected under unrestrained environment during signal acquisition. So, inorder to remove these artifacts, we have used Savitzky-Golay filter. Savitzky-Golayfilter is a smoothing filter that works by fitting consecutive subsets of contiguousdata points with a first-order polynomial, through linear least square (LLS) method.LLS is a significant type of statistical model known as linear regression, and it canbe calculated using Eq. 1:

ZPredicted = c + dY (1)

This equation represents least square regression line (LSRL) in which c is theintercept and d is the slope. The intercept c is computed through Eq. 2:


c = m(Z) − d (m(Y )) (2)

d = RSD (Z)

SD (Y)(3)

where m shows the mean, “SD” represents the standard deviation, and R representsthe correlation between Z and Y.

2.3 Segmentation

As the heart pumps, it produces S1, S2, S3, and S4 from which we will onlyconsider S1 and S2 because these sounds comprise dominant part of the PCG signal.Segmentation is performed to discriminate the systole and diastole periods based onS1 and S2 sounds. In order to segment S1 and S2 sounds, the proposed systemutilizes multiresolution wavelet decomposition as it gives best time and frequencyresolution of signal components by splitting low components into smaller bands. Itdecomposes the signal at nth level which is empirically selected as 2 in the proposedsystem.

We have selected Daubechies 2 as a mother wavelet because of its exceptionalproperties such as sharp filter transition band, orthogonality, fast computation,and time invariance. Different time and frequency components can be gener-ated by adding a scaling or a shifting factor to the mother wavelet as shownin Eq. 4 below:

∅(y) =∑∞

n=−∞aN∅ (Sy − n) (4)

Here “S” is the scaling factor and “n” is the shifting factor. In order to make thescaling function orthogonal, its functional area must be normalized as shown in Eq.5 below:

∫ ∞

−∞∅(y)∅ (y + 1) dy = δ0,1 (5)

In Eq. 4, aN represents number of coefficients, and its approximation and detailcoefficients for one of the randomly selected PCG signals are shown in Fig. 3.After extracting the S1 and S2 sounds, five distinct features are computed for theautomated identification of cardiovascular pathology. The first feature f1 is thesystole period that is computed by taking a mean of absolute difference betweenS2 and S1 intervals. The second feature f2 is the diastole period that is computed by

122 G. Sidra et al.

Fig. 3 Wavelet coefficients of randomly selected PCG signal

taking a mean of absolute difference between S1 and S2 intervals. The third featuref3 is the absolute difference between diastole and systole period as shown in Eq. 6below:

f3 = |f2 − f1| (6)

The fourth feature f4 is extracted through Shannon entropy, as expressed byEq. 7:

f4 = −∑m

j=1

∣∣Kpm,n

∣∣2log

∣∣Kpm,n

∣∣2(7)

where n = 0, 1, 2 . . . N, m =0, 1, 2, 3 . . . , 2N − 1. “m” signifies decompositionlevel, “n” denotes number of nodes for wavelet packet, and “p” represents the indexof scale. At last the fifth feature f5 is extracted by computing normalized energy asrepresented in Eq. 8:

f5 =∑m

j=1

∣∣Kpm,n

∣∣2(8)


Fig. 4 (a) Difference between S2 and S1 is systole, and S1 to next S2 is diastole periods in healthyPCG signal. (b) Difference between S2 and S1 is systole, and S1 to next S2 is diastole periods inabnormal PCG signal

Table 1 Extracted features

FeaturesType Cases F1 F2 F3 F4 F5

Normal Case 1 7121.4 6182.517 938.8828 1.6E + 11 7.85E + 09Case 2 24,737 61351.3 36614.33 5.4E + 11 2.34E + 10Case 3 132,990 69,851 63,139 5.3E + 11 2.26E + 10Case 4 381,217 65,349 315,868 3.7E + 12 1.45E + 11Case 5 13,268 15242.13 1974.125 1.1E + 11 5.57E + 09µ 111866.68 43595.0 83707.0 1.008E + 12 4.088E + 10S. D. 159123.10 30338.0 132350.0 1.5182E + 12 5.877E + 10

Abnormal Case 1 2110.85 1892.52 218.34 3.6E + 10 2.01E + 09Case 2 34,323 33934.5 388.5 1.8E + 11 8.16E + 09Case 3 4623.65 3916 707.65 9.8E + 10 5.02E + 09Case 4 3891.19 5610.28 1719.09 1.4E + 11 6.89E + 09Case 5 2557.23 2597.79 40.567 4.9E + 10 2.68E + 09µ 9501.18 9590.21 614.829 1.0060e + 11 5.7940E + 09S. D. 13912.40 13682.37 664.494 6.0620E + 10 2.3936E + 09

where “m” signifies decomposition level, “n” denotes number of nodes for waveletpacket, and “p” represents the index of scale.

Equations 7 and 8 represent the entropy and energy calculated from everysubband coefficient of wavelet packets (Fig. 4).

Table 1 shows feature vectors of arbitrarily selected five normal and abnormalPCG signals from PASCAL dataset.

124 G. Sidra et al.

2.4 Hybrid Classification

The proposed system employs hybrid classifier for the automated diagnosis ofCVD. Hybrid classification provides better performance as compared to individualclassifier. In classification stage, the 5D feature vector is fed to the supervised hybridclassifier which is formed through the ensemble of KNN, Naïve Bayes, and SVM.All of these classifiers were trained on 55 labeled samples from PASCAL dataset.The partial decision of each classifier is fused together through majority voting.

SVM

SVM is the first classifier used in the proposed system. It is one of the most popularand fast supervised classifier [18]. In the proposed system, it is implemented as anonlinear hyperplane by using Gaussian radial basis function (RBF) and multilayerperceptron kernel (MLP) as a kernel functions.

KNN

KNN is the second supervised classifier that has been used in the proposed system.It is a nonparametric classification algorithm that finds k-nearest neighbors of thetest sample using a distance metric, and then it assigns the class to the test samplewhich has majority votes among its “k” neighbors [19]. The proposed system usesEuclidean distance as a distance metric for KNN classifier where the best value of“k” is selected by measuring the maximum number of correctly classified samplesfor each value of “k.” The best selected “k” value was then used to classify thecandidate test sample.

Naïve Bayes

Naïve Bayes is the third supervised classifier in the proposed system. Naïve Bayesis a probabilistic nonlinear classifier that can classify correct samples by fitting anonlinear decision boundary. Since all the features in the proposed implementationhold the independence of events property, Naïve Bayes had a good performancein the proposed implementation. It is based on the Bayes’ rule as expressedin Eq. 9 [20]:

P (A|B) = P (B|A) ∗ P(A)

P(B)(9)


Table 2 Proposed system performance

Type Correctly classified Accuracy Sensitivity Specificity NPV

Abnormal 30/31 87.7% 96.7% 75% 94.7%Normal 18/24

f4

PCG Signal

Shannon Entropy

Wavelet Decomposition

Normalize Energy

f1 f3f2 f5

HybridClassifier

DatabaseCross

Validation Performance

Measurement

KNN

Naive Bayes

SVM

Fig. 5 Training phase of the proposed system

where P(B| A) is computed using univariate Gaussian distribution expressedby Eq. 10:

P (B|A) = 1

σ√

2πe

−(x−μ)2

2σ2 (10)

where μ is the mean and σ is the variance of the distribution. The decisions of allthree supervised classifiers are combined, and then a final decision is formed bymeasuring the majority votes between all three classifiers. After that, twofold cross-validation is performed between actual decisions and predicted decision, and theperformance of the hybrid classifier is measured which is also shown in Table 2.Figure 5 depicts the training phase of the proposed classification system.

126 G. Sidra et al.

We have measured the performance of the proposed system on a standardizedPASCAL dataset by computing sensitivity, specificity, accuracy, and negativepredictive value using Eqs. 11, 12, 13, and 14:

Sensitivity = TP

TP + FN

(11)

Specificity = TN

TN + FP

(12)

Accuracy = TP + TN

TP + TN + FP + FN

(13)

NPV = TN

TN + FN

(14)

Here, Tp represents true positives, TN represents true negatives, FP representsfalse positive, FN represents false negative, and NPV represents negative predictivevalues.

3 Results

The proposed dataset was tested on a publicly available PASCAL dataset thatcontained 55 PCG signals out of which 24 were normal and 31 were abnormalsamples. The dataset was annotated by multiple expert cardiologists. We havecross-verified the results of the proposed system against those annotations, andthe performance of the proposed system is shown in Table 2. The proposedsystem employs hybrid classification through the ensemble of SVM, KNN, andNaïve Bayes, and it takes around 3 s on average for the proposed system toautomatically predict the cardiac pathology on a machine with core i5 processorand 4GB DDR3 RAM. Apart from this, Fig. 6 shows four randomly selectednormal and abnormal PCG signals which are accurately classified by our proposedsystem.


Fig

.6C

lass

ified

PCG

sign

als:

(a)

norm

alsa

mpl

esan

d(b

)ab

norm

alsa

mpl

es.R

edco

lor

show

sth

ese

gmen

ted

hear

tbea

ts

128 G. Sidra et al.

4 Conclusions

This paper presents a fully automated system to diagnose cardiac pathologies byextracting heartbeats from PCG signals. The classification is based on the ensembleof SVM, KNN, and Naïve Bayes classifiers. The proposed method was applied ona total of 55 PCG signals in which 24 were normal signals and 31 were abnormalsignals. The proposed method was correctly classified 30/31 abnormal and 18/24normal PCG samples with accuracy, sensitivity, and specificity ratings of 87.7%,96.7%, and 75%, respectively. The proposed system was tuned in such a way to givemore priority in classifying abnormal samples correctly because it is more criticalnot to misclassify any diseased person as healthy. The proposed system can act asan aid to cardiologists to mass screen patients living in different geographical areasof the world. In future, this work can be extended by automatically identifying andgrading different pathological conditions of the heart.

Acknowledgment We are very thankful to PASCAL team for providing an online database ofannotated PCG signals.

References

1. WHO Eastern Mediterranean Region, Cardiovascular Diseases, World Health Organization.Retrieved from http://www.who.int/mediacentre/factsheets/fs317/en/ in September 2016.

2. CDC 24–7 Organization, Heart Disease Fact Sheet, Centers for Disease Control and Preven-tion. https://www.cdc.gov/dhdsp/data_statistics/fact_sheets/fs_heart_disease.htm in Septem-ber 2016.

3. United Health Foundation, Public Health Impact: Heart Disease, Americans Health Organiza-tion. http://www.americashealthrankings.org/explore/2015-annual-report/measure/CHD/state/ALL in September 2016.

4. Singh, M., & Cheema, A. (2013). Heart sounds classification using feature extraction ofPhonocardiography signal, 2013 International Journal of Computer Applications, pp. 13–17.

5. Bassiouni, M., Khalifa, W., El Dahshan, E. S. A., & Salam, A. B. M. (2015). A study on PCG asa biometric approach, 2015 IEEE Seventh International Conference on Intelligent Computingand Information Systems (ICICIS), Cairo, pp. 161–166.

6. Agrawal, K., Jha, A. K., Sharma, S., Kumar, A., & Chourasia, V. S. (2013) Wavelet subbanddependent thresholding for denoising of phonocardiographic signals, 2013 Signal Processing:Algorithms, Architectures, Arrangements, and Applications (SPA), Poznan, pp. 158–162.

7. Shankar, N., & Sangeetha, M.S. (2013). Analysis of Phonocardiogram for detection of cardiacmurmurs using wavelet transform. 2013 International Journal of Advanced Scientific andTechnical Research, pp. 350–357.

8. Bhogawar, S. (2014) Cardiac sound analyzer and monitor, 2014 International Journal ofResearch in Engineering and Applied Sciences (IJREAS).

9. Sankar, D. S. V., & ARoy, L. P. (2014). Principal component analysis (PCA) approachto segment primary components from pathological phonocardiogram, 2014 InternationalConference on Communication and Signal Processing, Melmaruvathur, pp. 910–914.

10. Vernekar, S., Nair, S., Vijaysenan, D., & Ranjan, R. (2016). A novel approach for classificationof normal/ abnormal phonocardiogram recordings using temporal signal analysis and machinelearning, 2016 Computing in Cardiology Conference (CinC), Vancouver, BC, pp. 1141–1144.

http://www.who.int/mediacentre/factsheets/fs317/en

https://www.cdc.gov/dhdsp/data_statistics/fact_sheets/fs_heart_disease.htm

http://www.americashealthrankings.org/explore/2015-annual-report/measure/CHD/state/ALL


11. Springer, D. B., Tarassenko, L., & Clifford, G. D. (2014). Support vector machine hidden semi-Markov model-based heart sound segmentation, Computing in Cardiology 2014, Cambridge,MA, pp. 625–628.

12. Pedrosa, J., Castro, A., & Vinhoza, T. T. V. (2014). Automatic heart sound segmentation andmurmur detection in pediatric phonocardiograms, 2014 36th Annual International Conferenceof the IEEE Engineering in Medicine and Biology Society, Chicago, IL, pp. 2294–2297.

13. Langley, P., & Murray, A. (2016). Abnormal heart sounds detected from short duration unseg-mented phonocardiograms by wavelet entropy, 2016 Computing in Cardiology Conference(CinC), Vancouver, BC, pp. 545–548.

14. Suhas, K., Kumar, R. H., Nayak, S. H., & Krupa, B. N. (2016). A hybrid model for recog-nizing cardiac murmurs from phonocardiogram signal, 2016 IEEE Annual India Conference(INDICON), Bangalore, pp. 1–6.

15. Yazdani, S., Schlatter, S., Atyabi, S. A., & Vesin, J. M. (2016). Identification of AbnormalHeart Sounds, 2016 Computing in Cardiology, pp. 1157–1160.

16. Springer, D. B., Brennan, T., Ntusi, N., Abdelrahman, H. Y., Zuhlke, L. J., Mayosi, B. M.,Tarassenko, L., & Clifford, G. D. (2016). Automated signal quality assessment of mobilephone-recorded heart sound signals. Journal of Medical Engineering and Technology, 40(7–8), 342–355.

17. Bentley, P., Nordehn, G., Coimbra, M., Mannor, S., & Getz, R. Classifying Heart Sounds Chal-lenge, Sponsored by PASCAL. Retrieved from: http://www.peterjbentley.com/heartchallenge/in September 2016.

18. Fan, R. E., Chang, K.-W., Hsieh, C.-J., Wang, X.-R., & Lin, C.-J. (2008). LIBLINEAR: Alibrary for large linear classification. Journal of Machine Learning Research, 9.

19. Altman, N. S. (1992). An introduction to kernel and nearest-neighbornonparametric regression. The American Statistician, 46(3), 175–185.https://doi.org/10.1080/00031305.1992.10475879.

20. Narasimha Murty, M., & Susheela Devi, V. (2011). Pattern recognition: An algorithmicapproach. Bangalore: Springer ISBN 0857294946.

http://www.peterjbentley.com/heartchallenge

http://dx.doi.org/10.1080/00031305.1992.10475879

Modeling and Simulationof Resource-Constrained VaccinationStrategies and Epidemic Outbreaks

Rehan Ashraf, Bushra Zafar, Sohail Jabbar, Mudassar Ahmad,and Syed Hassan Ahmed

1 Introduction

Historical evidences indicate that infectious diseases have claimed more lives thanarmed conflicts and natural disasters put together [1]. Infectious diseases whichcan be cured by vaccination are termed as vaccine-preventable diseases, and deathresulting from such diseases is referred to as vaccine-preventable death. For somediseases vaccines are available and for some others are in pipeline process [2]. Low-income countries currently have a relatively higher share of deaths and are at highrisk of epidemics compared with high- and middle-income countries [3]. Over-crowded cities, unsafe drinking water, inadequate sanitation, poor socioeconomicconditions, low health awareness, and inadequate vaccination coverage are some ofthe factors that make these countries more prone to viral attacks. Left unchecked, itcould be disastrous. It poses a threat to developed countries as well.

Globally, influenza epidemics result in about 3 to 5 million cases of severe illness,and about 250,000 to 500,000 fatal cases lead to death [4]. Smallpox, Ebola, yellowfever, cholera, and polio are listed as the deadliest epidemics and still exist as apotential threat [5]. Emerging and reemerging diseases are posing threats to humanlives. More and more people are being exposed to diseases like mumps, measles, andrubella, which were once considered to be vanished. Some vaccines are given as aseries of shots instead of a single dose. And some require periodic booster shots [6].

This paper models a complex emerging phenomenon whereby micro-level agentsgenerate macro-level behavior following simple rules. Theadvantage of modeling is

R. Ashraf · B. Zafar · S. Jabbar · M. Ahmad (�) · S. H AhmedDepartment of Computer Science, National Textile University, Faisalabad, Pakistane-mail: [email protected]; [email protected]; [email protected]


131





https://doi.org/10.1007/978-3-319-96139-2_13

132 R. Ashraf et al.

that we are able to simulate conditions which are not practicable to visualize in thereal world. In addition one can ascertain parameter values that can lead to a deadend. The model follows the simple SIR model state transitions with demographics.There are various factors which influence the survival of a directly transmittedvirus in a population [7]. Some of these factors were illustrated in a model byUri Wilensky [8, 9] in NetLogo. The interaction of these factors leads to differentoutcomes and behaviors depicting how virus will propagate in population. Thefactors which might affect the survival of a directly transmitted virus in a populationare:

– Population density: Population density determines how frequently the infected(I), susceptible (S), and immune (R) agents approach each other. In areas of densepopulation, contact rate will be high as compared to sparsely populated areas.

– Population turnover: New individuals will be reproduced as existing die ofinfection or old age. In this model it is assumed that an individual can reproduceon average four times during his life span and only healthy agent can reproduce.The new agents born will be all healthy and susceptible.

– Immunity: Once a person gets infected, he can either recover or die. It mightlast a lifetime or wears off with time. In this model immunity is assumed to beeternal. Once attained, it will last forever.

– Infectiousness: It determines the ease with which the virus spreads. High valueof infectiousness indicates that the disease is likely to spread easily. Its lowvalue indicates fewer chances of infection. Some viruses spread easily as themicroparasite mycobacterium tuberculosis that causes TB, while others requiresignificant contact as in the case of HIV that is responsible for AIDS.

– Duration: It determines how long the infected agents will survive before theyare either cured or die. It is also determined by user. It basically opens upopportunities for virus to be transmitted to other hosts.

The original virus model simulates the transmission and perpetuation of a virusin a human population [8]. But it does not include any option to assess the impactof any health interventions to contain epidemic spread. To analyze the impact ofvaccination strategies, the following extensions were made to the original model:

– Disease can be reintroduced in population dynamically at any time.– Vaccination strategy random and fixed to vaccinate the population. In case of

random vaccination, a random subset of population is selected at each scheduledinterval. In fixed population vaccination, a particular subset is vaccinated aftereach interval.

– Vaccine efficacy period as it varies across diseases. For some diseases therecommended interval for next dose is in months, and for others it is in years.

– Dosage refers to the total no of recommended doses that the individual mustacquire in order to attain immunity.

Modeling and Simulation of Resource-Constrained Vaccination Strategies and. . . 133

We have assumed an open population contrary to closed population. In closedpopulation no births, deaths, or travel takes place, whereas in open population birthsand deaths take place, and people can move into or out of the population [10]. Openpopulation is more realistic approach to explain the spread of endemic diseases indeveloping countries with high birth rates [11]. Also there is homogeneous mixingof population, i.e., every agent has equal chance of interacting with another andcontracting the disease. Different patterns of emergence were identified from theoriginal model, and impact of vaccination strategy was analyzed on those patterns.

It can be further extended by considering the other demographic characteristics ofpopulation- and disease-specific parameters. This model is a general abstraction andcan be applied to different scenario of an epidemic outbreak. The principal resultsof the study were that fixed population vaccination was found to be more effectivecompared to random population vaccination.

2 Related Work

A lot of research work is being done to optimize the vaccination strategies tominimize the impact of infectious disease outbreak [12–14]. Alexander et al.proposed a mathematical model to determine whether the approach of givingbooster shots would help to eradicate the disease from population. They conductedexperiments considering measles epidemic, and their findings suggest that givingbooster shots might be helpful to control the spread but not necessarily eradicateit [15].

Laskowski et al. analyzed the impact of early vaccination strategies in case ofinduction of single-dose or two-dose vaccine protection. They observed that in caseof highly infectious disease, vaccine delivery had no substantial impact after onsetof epidemic [16].

Vaccines help to attain herd immunity. If some portion of population is vac-cinated, they serve as barriers between the sick and healthy. It leads to eithereradication or containment of disease. This phenomenon is termed as herd immu-nity. Vaccines helped eradication of epidemic diseases like polio, smallpox, measles,whooping cough and many others. One of the reasons of reemergence of epidemicdiseases is the ignorance of people to get vaccinated, and one of the most concernsabstaining people are the adverse effects [17].

CDC recommended periodic vaccination for 17 vaccine preventable diseases,though the recommendation is for the USA and may vary regionally depending onepidemic dynamics [18]. The dose of vaccine required to develop immunity againstspecific disease varies in case of different vaccines (effective against particular typeof infection). Sometimes single dose of vaccine is effective for life like in the case ofMMR (measles, mumps, and rubella) vaccine, while in some cases, a booster dose(administration of vaccine after primary dose) is either required to maintain antibody


level against target antigen or after a specified period to remind the defense/immunesystem to fight against the noxious agent.

3 Proposed Model

The original virus model simulates the persistence and perpetuation of a micropar-asitic pathogen in a human population. Persistence means how long virus willsurvive in populations which can be through generations. It does not account for anycountermeasures that can lead to eradication of disease. Eradication means to helpwipe out the disease from population. Extinction refers to obliteration of population.The world is split into equally sized patches which is a square piece of “ground.”Agents move over patches. A turtle can on average produce four offsprings inlifetime. Newborns are all initially in healthy state. We have set an average life spanof agent to 50 years, and each tick corresponds to 1 week. The agents can freelymove inside the world.

The model follows the simple SIR model state transition rules, with demograph-ics in a bounded population of fixed size. In this model the world is confined to a sizeto allow for maximum agents, but the population size is allowed to vary with timesubject to births and deaths. The agents can be in one of three states and undergostate transitions. Initially all agents are in a healthy state, and a few are in the infectedstate.

– Healthy but susceptible (S) to infection. The S agents can acquire disease ifcomes across a sick agent. And the probability to acquire disease is determinedby infectiousness of disease.

– Agents who are in sick and infectious (I) state are those that have acquired diseaseand can infect others. These agents can either recover or ultimately die, and theprobability to recover is determined by chance-recover option.

– In the original model, healthy and immune (R) agents are those who haverecovered from disease. We have extended the definition of immune agents asthose who have either recovered or have completed their recommended courseof vaccination. They are resilient to infection.

At each epoch the agents move randomly and depict some behavior. Theproperties and behaviors of agents at micro level of analysis are summarized inTable 1. The prescribed behaviors are executed repeatedly until either the time frameset as 50 years in this model is completed or the population is extinct, whicheveris earlier. Table 2 summarizes the properties and behaviors of environment. AndTable 3 explains the modified model.

The original model is referred to as NV acc for no vaccination. To vaccinate afixed set of population at periodic intervals is termed as fixed population vaccination(FP V acc). And to vaccinate a random set of population is denoted as RP V acc.The interface is shown in Fig. 1. RP V acc models the behavior of people in realworld, as ignorance to complete course.


Table 1 The agent properties, behaviors, and description

Properties and behaviors Description

Life span The average life span of an agent

Age Holds age of agents in weeks

Health-status Sick? If true, the agent is infectious(I)

Immune? If true, the agent cannot be infected (R)

Sick-time Holds the time since the agent is infectious (in weeks)

Average-offspring Average number of offspring’s of an agent

Chance-recover Controls the likelihood that an infection will weaken, before theyeither recover or die

Get-older Agents die if their age exceeds the set life span

Move Agent moves about at random

Infect I agents infect S agents on same path

Reproduce If agents are less than carrying, healthy agents can reproduce

Recover If agents have survived past duration, they either recover or die

Table 2 The environment properties and behaviors of agents

Properties and behaviors Description

People The initial number of people in the world

% infected Holds percentage of infected agents in population

% immune Holds percentage of immune agents in population

Carrying-capacity The number of agents that can be in world at on time

Setup-constants Some important parameters are set as constants which can beexposed as sliders

Setup-turtles Creates a variable number of agents of which ten are infected

Update-global-variables Agents counting variables are updated

Table 3 The modified model

Modification Original Proposed

Properties – Vacc-r: No. of vaccine dosages an agent receivedsince birth

– Vacc-no: Total No. of vaccine dosages delivered sincebirth of agent

– Dose: Recommended No. of vaccine dosages for anagent

– V_EP: Vaccine efficacy period between successivedoses

Behavior – N_Vacc FP_Vacc: For vaccination of a fixed subset of popu-lation

– Introduce-disease: To break in disease at any time inpopulation with ten infected agents

Update-global-variables

The original behavior kept count in real time. It wasmodified to keep track of I , R, and

S agents with every 5 years interval


Fig. 1 Flow-chart for vaccination module

It is assumed that agents who do not follow schedule properly have a tendencyto acquire disease. Probability is set so if an agent has received less than 80%of recommended dosages at a particular time, he is susceptible to infection. Thenumber of vaccine dosages per lifetime of individual can be set by the user it variesacross diseases.

Vaccine efficacy period refers to the time period for which the vaccine inducesimmunity in agent. Immunity can be lifelong, or for some disease it wears off withtime. A simple control flow for vaccination module is shown in Fig. 2. It doesnot represent the flow of entire simulation, just the concerned module. NV acc isconsidered the setting with no vaccination. Depending on FP V acc or RP V acc

choice, the population is divided into equivalent subsets, respectively.V acc − r refers to vaccination dosages received by the agent, and vacc − no is

the number of schedules conducted since his birth which gets incremented in everycampaign even if the agent is not selected for vaccination. Vaccine efficacy periodrefers to time for which vaccine induced protection lasts. So if the selected subsethas not received complete course and the efficacy period expired, it gets vaccinatedand is checked for immunity criteria. The agents who meet the set immunity criteriaacquire immunity which for this model is assumed eternal. The process is repeatedfor all agents until the vaccination number equals dosage. If vacc − no exceedsdosage, the individual is not considered eligible for vaccination.


Fig. 2 Emergent pattern healthy and immune agents

4 Simulation Results

Firstly the macro-level behavior of original model was analyzed for certain parame-ter space using parameter sweeping. The impact of FP-Vacc and RP-Vacc was alsoobserved for the same parameter space, and the impact on emergent patterns wasanalyzed. The experimental design is shown in Table 4.

We have broadly categorized infectious diseases in these experimental settings.The output of behavior space was analyzed using R statistical tool. The diseasesspecified in table are obtained from the WHO list for available vaccines. The aboveexperiment space was designed by observations made from parameter sweeping andalso from the analysis of a related simulation study [19]. We have selected dose 4as an average case, as is recommended for diseases like polio, hepatitis B, and DTP.V − EP also varies across diseases from weeks to months to years, so we haveselected the interval as 1 year for current experimental settings.


Table 4 Experiment design

Exp. No. I C-R D P Dose V-EP Opt Diseases

1 5 0 99 150 4 1Y N-Vacc, FP-Vacc, RP-Vacc AIDS

2 99 0 2 150 4 1Y N-Vacc, FP-Vacc, RP-Vacc Diphtheria, measles

3 99 0 99 150 4 1Y N-Vacc, FP-Vacc, RP-Vacc Influenza, polio

Fig. 3 Bounded pattern healthy and immune agents

The results of experiment 1 are shown in Fig. 3. Here it is quite obvious that in thecase of FP V acc, we have a high density of healthy and immune agents compared toRP V acc. RP V acc has high density of infected agents. NV acc leads to extinctionof population in these settings. It is observed that disease with low infectiousnessand long duration survive in population for long.

The results of 2nd experimental settings are shown in Fig. 4. Again it is evidentthat FP V acc appears to be more efficient here. In the case of FP V acc, the diseasewas almost eradicated from population, while in RP V acc a pattern of persistenceis observed. A high level of healthy agents in RP V acc corresponds to the fact as itis bounded population; the new agents born to replace dead are healthy. RP V acc


Fig. 4 Comparison performance of FP_Vacc, RP_Vacc, and N_Vacc

could be the possible reason of persistence of endemic diseases like TB and measlesin developing countries.

Experiment 3 settings also affirms to the efficacy of FP V acc as can be seen inFig. 5. There is no significant difference in outcomes of NV acc and RP V acc. Thehigh density of healthy agents in FP V acc corresponds to herd immunity whereimmune agents serve as barriers between sick and infected and add to overall healthof population.

The study was focused to observe optimal vaccination strategy in the case ofscarce resources, so FP V acc prove to be better though not the optimal solution.However, it would be interesting to observe the outcomes by implying otherinterventions in addition to vaccination. RP V acc results do not differ much fromNV acc.

The results indicate that if some significant portion of the population realizesits responsibility and takes complete course of vaccination, they develop immunityand serve as barriers between sick and healthy and hence contribute to the health ofpopulation, whereas if individuals do not complete the course of vaccination as isnormally observed in routine life, that would significantly lower the immunity andresult in outbreaks.


Fig. 5 User Interface of the simulation environment

5 Conclusion

The principal results of the study were that fixed population vaccination was foundto be more effective compared to random population vaccination. This can help toexplicate why some diseases are not eradicated and reemerge in spite of aggressivevaccination campaigns. If vaccination coverage of whole population is not feasible,even a fixed subset of population can help to achieve herd immunity and may leadto eradication of disease depending on disease-specific traits. We used people toepitomize the population in the model, but it can be adapted to the livestock animalsas well. This model simulates directly transmissible diseases; we plan to extendthis further to model vector-borne transmission as well. Moreover, agents in thismodel move randomly and are free to interact with others. Social networks can beintroduced to extend this model. We intend to extend this work further by applying itto infectious disease scenarios using more precise parameters, considering disease-specific and other demographic characteristics of populations.


References

1. Inhorn, M. C., & Brown, P. J. (1990). The anthropology of infectious disease. Annual Reviewof Anthropology, 19, 89–117.

2. WHO. Immunization, Vaccines and Biologicals. Available: http://www.who.int/immunization/diseases/en/. Accessed: 20 Nov 2015.

3. WHO. (2008). Global burden of disease 2004 update: Selected figures and tables. Geneva:World Health Organization.

4. Cohen, J., & Enserink, M. (2002). Rough-and-tumble behind Bush’s smallpox policy. Science,298, 2312–2316.

5. Feldmann, H., Jones, S., Klenk, H.-D., & Schnittler, H.-J. (2003). Ebola virus: From discoveryto vaccine. Nature Reviews Immunology, 3, 677–685

6. Immunizations. Available: http://kidshealth.org/teen/school_jobs/college/immunizations.html.Accessed: 25 Dec 2015.

7. Yorke, J. A., Nathanson, N., Pianigiani, G., & Martin, J. (1979). Seasonality and therequirements for perpetuation and eradication of viruses in populations. American Journal ofEpidemiology, 109, 103–123.

8. Wilensky, U. (1998). NetLogo Virus model. Available: Version: 5.1.0. http://ccl.northwestern.edu/netlogo/models/Virus.

9. Wilensky, U. (1999). Netlogo. Available: Version: 5.1.0. http://ccl.northwestern.edu/netlogo/.10. Rothman, K. J., Greenland, S., & Lash, T. L. (2008). Modern epidemiology. Philadelphia:

Lippincott Williams & Wilkins.11. Allen, L. J., Brauer, F., Van den Driessche, P., & Wu, J. (2008). Mathematical epidemiology.

Berlin/Heidelberg: Springer.12. Holzmann, H., Hengel, H., Tenbusch, M., Doerr, H. W. (2016). Eradication of measles:

Remaining challenges. Medical Microbiology and Immunology, 205, 201–208.13. Osterholm, M., Moore, K., Ostrowsky, J., Kimball Baker, K., Farrar, J., & Wellcome Trust-

CIDRAP Ebola Vaccine Team. (2016). The Ebola Vaccine Team B: A model for promoting therapid development of medical countermeasures for emerging infectious disease threats. TheLancet Infectious Diseases, 16, e1–e9.

14. Prada, J., Metcalf, C., Takahashi, S., Lessler, J., Tatem, A., & Ferrari, M. (2017). Demo-graphics, epidemiology and the impact of vaccination campaigns in a measles-free world Canelimination be maintained? Vaccine, 35, 1488–1493.

15. Alexander, M., Moghadas, S., Rohani, P., Summers, A. (2006). Modelling the effect of abooster vaccination on disease epidemiology. Journal of Mathematical Biology, 52, 290–306.

16. Laskowski, M., Xiao, Y., Charland, N., Moghadas, S. (2015). Strategies for early vaccinationduring novel influenza outbreaks. Scientific Reports, 5, 1–13.

17. Libster, R. (2014). The power of herd immunity. Available: https://www.ted.com/talks/romina-libster-the-power-of-herd-immunity/transcript?language=enTEDxRiodelaPlata 14:41 FilmedNov 2014

18. Atkinson, W. L., Pickering, L. K., Schwartz, B., Weniger, B. G., Iskander, John K., &Watson, J. C. (2002). General recommendations on immunization. Recommendations of theAdvisory Committee on Immunization Practices (ACIP) and the American Academy of FamilyPhysicians (AAFP) MMWR Recomm Rep, 51, 1–35.

19. Johnson, M., Gilliam, T., & Karsai, I. (2009). The effect of infectiousness, duration of sickness,and chance of recovery on a population: A simulation study. Bios, 80, 99–104.

http://www.who.int/immunization/diseases/en/

http://www.who.int/immunization/diseases/en/

http://kidshealth.org/teen/school_jobs/college/immunizations.html

http://ccl.northwestern.edu/netlogo/models/Virus

http://ccl.northwestern.edu/netlogo/models/Virus

http://ccl.northwestern.edu/netlogo/

https://www.ted.com/talks/romina-libster-the-power -of-herd-immunity/transcript?language=en TEDxRiodelaPlata

https://www.ted.com/talks/romina-libster-the-power -of-herd-immunity/transcript?language=en TEDxRiodelaPlata

Big Data in Healthcare: A Survey

Muhammad Mashab Farooqi, Munam Ali Shah, Abdul Wahid,Adnan Akhunzada, Faheem Khan, Noor ul Amin, and Ihsan Ali

1 Introduction

The collection of very large and complex set of raw facts and figures is calleddata; it is very difficult to process using ordinary database management systems[1]. We need certain tools, techniques and procedures to create, manipulate andmanage very large set of Big Data. Big Data does not have just large size,but it also contains complex, heterogeneous, noisy, longitudinal and voluminousdata [2]. The challenges which the organization faces in handling the Big Dataare capturing, searching, storing and analysing the data. The five dimensions ofBig Data (5 V’s) are volume, variety, velocity, veracity and value [3, 4]. Themain reason for the growth of the complexity and abundance of data is that themedical practice is moving to evidence-based healthcare. The other reason for theabundance and increased complexity of the Big Data is the development of the newtechnologies and tool such as mobile application capturing devices and sensors,as these devices collect huge amount of data which is to be stored somewhere.Big Data leads towards the increased patient social communication [5], to makeonline appointment, to search patient’s record online and to check the availability

M. M. Farooqi · M. A. Shah · A. Wahid · A. Akhunzada (�)Department of Computer Science, COMSATS Institute of Information Technology, Islamabad,Pakistane-mail: [email protected]; [email protected]

F. Khan · N. ul AminDepartment of Computer Science, Bacha Khan University Charsadda, Charsadda, Pakistan

I. AliDepartment of Computer Systems and Technology, Faculty of Computer Science and InformationTechnology, University of Malaya, Kuala Lumpur, Malaysiae-mail: [email protected]


143





https://doi.org/10.1007/978-3-319-96139-2_14

144 M. M. Farooqi et al.

of the doctor online. Handling a large amount of data efficiently is a challengefor the organization. Traditional host or service in Big Data is shifting towardsdata-centric architecture and model [6]. Components included in the Big DataArchitecture Framework are Big Data infrastructure, data structures and models,Big Data analytics, Big Data lifecycle management and Big Data security [7, 8].

The technologies such as cloud computing are very necessary for the providenceof a stage for the computerization of all processes in data collection, storing,processing and visualization. By ensuring that the patient gets the most effectivetreatment, better care quality and efficiency can be achieved in the healthcare [2, 9].However, these new technologies not only made the data huge and complex but alsomade it difficult to handle and process because of its unstructured nature. Moreover,data is recorded from different devices like sensors and smartphones within a shortperiod of time. These data are stored in different formats which can be regarded asa challenge in Big Data. In [9] Health-CPS built on cloud and Big Data analytictechnologies are used by some health organizations to provide a more convenientservice and environment of healthcare. In order to become responsive, healthcareorganizations need to be agile. They can achieve agility by optimizing theiroperations. Service-oriented architecture (SOA) and business process management(BPM) are adopted by some healthcare organizations which can provide flexible,dynamic and cloud-ready infrastructure [10]. Big Data analytics helps to understandlarge amount of data and to categorize it, then predict the outcomes before it happensand suggest the treatments [11].

This paper presents the review of the characteristics of Big Data, tools andtechniques, challenges and limitations and architectures used by the healthcareorganization.

2 Characteristics of Big Data in Healthcare

Many researchers have studied and worked on Big Data in healthcare; there aremany challenges, prospects and resolutions in Big Data in healthcare. There aresome characteristics of Big Data known as dimensions of the Big Data. All thedimensions are discussed briefly in this section.

Big Data is characterized as 5 V’s by many authors [3, 4, 12]. These character-istics are volume, variety, velocity, variability and veracity. As the time is passing,these dimensions are increasing; now in 2017 we have 42 V’s [13]. 42 V’s arementioned in this paper and summarized in Table 1.

Table 1 discusses the 42 V’s of the Big Data, and the dimensions will continueto increase as the Big Data develops further. The generic concept of Big Data isportrayed in Fig. 1.

Big Data in Healthcare: A Survey 145

Tab

le1

Cha

ract

eris

tics

ofB

igD

ata:

the

42V

’s

Cha

ract

eris

tics

Mea

ning

Cha

ract

eris

tics

Mea

ning

Vol

ume

Size

ofda

tage

nera

ted

inhe

alth

care

Ven

ueN

ocu

stom

erw

orks

tati

onin

clou

d(l

ogic

ally

)V

eloc

ity

The

spee

dat

whi

chda

tais

gene

rate

dV

ersi

onco

ntro

lY

ouar

eus

ing

itri

ght?

Var

iabi

lity

The

chan

gein

data

sem

anti

cs,d

ata

stru

ctur

e,da

tafo

rmat

and

data

rate

.Evo

lvin

gbe

havi

our

inda

taso

urce

Vex

edPo

tent

ialo

fda

tasc

ienc

eto

hand

leco

mpl

icat

edpr

oble

ms

Var

iety

Div

erse

sour

ceof

data

from

whi

chdi

ffer

ent

sour

ces

data

ispr

oduc

edV

ibra

ntPr

ovis

ion

ofin

sigh

tby

data

scie

nce

Ver

acit

yT

hequ

ality

ofda

tath

atis

prod

uced

,dat

aac

cura

cyV

iral

How

fast

data

spre

ads

Val

ueU

sefu

ldat

aV

irtu

osit

yC

raze

toge

tmor

ekn

owle

dge

abou

tBig

Dat

aV

alid

ity

Dat

aqu

alit

y,go

vern

ance

,dat

am

anag

emen

ton

mas

sive

Vis

uali

zati

onT

hew

aycu

stom

erin

tera

cts

wit

hm

odel

sV

arni

shIn

tera

ctio

nof

end-

user

sw

ith

our

wor

km

atte

rs,a

ndpo

lish

coun

tsV

ogue

Art

ifici

alin

tell

igen

cew

illb

ecom

e?

Vau

ltIm

port

ance

ofda

tase

curi

tyV

oodo

oD

eliv

erre

sult

sw

ith

real

-wor

ldim

pact

Vei

lE

xam

ine

late

ntva

riab

les

from

behi

ndth

ecu

rtai

nV

ulpi

neD

ata

lead

sto

ane

wte

chno

logy

Ver

dict

Peop

leaf

fect

edby

mod

el’s

deci

sion

Vag

uene

ssC

onfu

sion

abou

tmea

ning

ofB

igD

ata

and

tool

sth

atar

eus

edV

erse

dK

now

ledg

eab

out:

mat

hem

atic

s,st

atis

tics

,pro

gram

min

g,et

c.V

alou

rTa

ckle

the

big

prob

lem

infa

ceof

Big

Dat

aV

etV

etti

ngth

eas

sum

ptio

nsw

ith

evid

ence

Van

eU

ncle

ardi

rect

ion

ofde

cisi

on-m

akin

gV

iabi

lity

Dif

ficul

tto

buil

dro

bust

mod

els

Van

illa

Sim

ple

met

hods

ifta

ckle

dw

ith

care

,can

prov

ide

valu

eV

ictu

alB

igD

ata

fuel

ofda

tasc

ienc

eV

anta

gePr

ivil

eged

view

ofco

mpl

exsy

stem

sV

isib

ilit

yD

ata

scie

nce

prov

ides

visi

bili

tyin

toco

mpl

exda

tapr

oble

ms

Var

min

tA

sda

tage

tsbi

gger

,so

doso

ftw

are

bugs

Viv

ify

Abi

lity

ofda

tasc

ienc

eto

cope

wit

hev

ery

real

-lif

eas

pect

Vas

tnes

sB

igne

ssof

Big

Dat

aV

oice

Abi

lity

tosp

eak

wit

hkn

owle

dge

Vat

icin

atio

nA

bili

tyto

fore

cast

Vol

atil

ity

Dat

ash

ould

neve

rbe

mis

sing

Vee

rC

hang

edi

rect

ion

acco

rdin

gto

cust

omer

need

Voy

age

Incr

easi

ngkn

owle

dge

Ven

ueD

istr

ibut

edhe

tero

gene

ous

data

from

mul

tipl

epl

atfo

rms

Var

ifoc

alIt

isab

outt

rees

and

fore

stV

ocab

ular

yD

ata

mod

els,

sem

anti

csth

atde

scri

beth

eda

ta


Volume

Velocity

Variety

Veracity

Value

Validity

Variability

Venue

Vocabulary

BigData

Vagueness

Size of Data

Distributed HeterogeneousData from Multiple Platforms

Data Models, Semantics thatdescribes data Structure

Confusion over Meaning of BigDataand Tools used

Dynamic, Evolving Behaviorin Data Source

The Speed at which Datais Generated

Different type of Data

Data Accuracy

Useful Data

Data quality, Governance Moster DataManagement on Massive

Fig. 1 Big Data concept

3 Tools and Techniques for Analysing Big Data

With the advancements in the information and communication, the healthcare datahas exceeded from exabyte to petabytes which is increasing gradually [4]. Withthis growth rate, it is difficult to handle large amount of data using traditional dataarchitectures and models. Health organizations have two options: whether they canuse open source or the commercial solutions available [14]. Some of the refinedtools, technologies and platforms for analysing Big Data are discussed in Table 2.

4 Challenges and Limitations

The emerging field of Big Data poses many challenges, limitations and issues as thehealthcare data is increasing [15]. As the industry is facilitating with the advantagesof Big Data, security and privacy issues are the main problem arising as the threatsand vulnerabilities keep on increasing [16]. Some of these challenges are discussedby the researchers which are mentioned in this section. Figure 2 shows Big Data inhealth cloud.


Tab

le2

Com

para

tive

anal

ysis

ofto

ols

Tool

sPl

atfo

rms

Type

ofda

taba

ses

Adv

anta

ges

Lim

itat

ions

Goo

gle

Big

Que

ryC

loud

-bas

edan

dop

en-s

ourc

epl

atfo

rmC

olum

nar

data

base

Rep

lica

teda

taac

ross

vari

edda

tace

ntre

sIn

dexe

sar

eno

tsup

port

edby

Goo

gle

Big

Que

ry

Had

oop

Clo

ud-b

ased

and

open

-sou

rce

plat

form

Itis

ano

n-re

lati

onal

data

base

Stor

esda

taan

dst

ruct

ure

No

secu

rity

and

tech

nica

lsup

port

Red

uce

Clo

ud-b

ased

and

open

-sou

rce

plat

form

Itis

ano

n-re

lati

onal

data

base

Itsu

ppor

tsbo

thun

stru

ctur

edan

dse

mi-

stru

ctur

edda

taN

oin

dexi

ngca

pabi

lity

Jaql

Itis

apr

opri

etar

yqu

ery

lang

uage

Itis

aqu

ery

lang

uage

for

Java

Scri

ptob

ject

nota

tion

Itsu

ppor

tsbo

thst

ruct

ured

and

sem

i-st

ruct

ured

data

The

reis

nocl

ient

char

acte

rize

dso

rts;

this

sugg

ests

patt

ern

data

isju

sta

requ

irem

ent

onco

ncei

vabl

ees

tim

atio

nsof

ado

mai

nM

icro

soft

Win

dow

sA

zure

Clo

ud-b

ased

and

open

-sou

rce

plat

form

Itis

are

lati

onal

data

base

Aga

inst

stru

ctur

ed,u

nstr

uctu

red

and

sem

i-st

ruct

ured

,itc

anm

ake

rela

tion

alqu

erie

s

Imm

ense

data

base

sar

eun

real

isti

con

Win

dow

sA

zure

asit

has

rest

rict

edsi

ze


Fig. 2 Big Data healthcare cloud

4.1 Data Governance

Regulating and managing of data is data governance. As healthcare industry ismoving towards the healthcare analytics, data governance is a big challenge [16].The data generated in healthcare is assorted in nature, and it requires standardizationand governance.

4.2 Security Analytics

Predicting and analysing security threats are of the highest need in the growinghealthcare industry. Healthcare industry suffered from these types of security attacksranging from stealthy malware to distributed denial-of-service (DDoS) attack.Social engineering attacks are also on the rise, and they are very difficult to predict.As healthcare industry is depending more on medicinal services innovations to settleon better educated choices, security investigation will be the primary concentrationof any plan for the cloud-based SaaS arrangement facilitating protected healthinformation (PHI) [16]. Healthcare IT providers can remove dangers and threatsin real time and can emit them before they impact the social insurance framework.


4.3 Ethical and Moral Challenges

Moral test incorporates information protection, control of access to patients’ data,secrecy and viable trade of data between the patients. Because of increment ininformation volume, the mix of information from various sources turns into achallenge. The entrance of patient’s data in a fitting way turns into an issue. Contextsensitivity is one of the moral tests that Big Data in healthcare is confrontingnowadays. It is imperative to guarantee the context sensitivity. Context of Big Datacontrasts in huge ways from different sorts of Big Data exercises. A context delicatecomprehension may reveal that a few information may not be reasonable insidecorporate movement [17].

4.4 Security and Privacy Issues

Security and privacy issues are another challenge facing the integration of diversesource of healthcare information. Healthcare data and mobile ad hoc are opento security threats like inappropriate access of the patients’ data and unapprovedutilization of patient data. Hence, healthcare providers are facing security andprivacy issues, and they fear to share health information using electronic medicinalservices frameworks. Privacy deals with the legal and ethical restrictions, and aquestion arises on which piracy issue must be taken care of [18]. The privacy policyis all about the implementation of procedures with respect to authorization andpermission of specific functions [19].

5 Discussion on Big Data Architecture

Big Data analytics assumes an essential part in anticipating the crisis circumstancebefore it happens. Big Data analytics uses Hadoop [1] for the real-time investigationon the enormous amount of data. Data analytics is essential because without legit-imate information investigation techniques, such information is useless. Hadoop isan open-source Apache Software Foundation project which is composed in Java.It has two main components, HDFS and MapReduce programming framework,which are closely related to each other [20]. In this approach, cost is reducedby the effective analysation of the data. With the assistance of machine learningalgorithm, data patterns and relationship between them are analysed which helpsin making valuable decisions. The author worked on Big Data generation, datacharacteristics, security concerns in Big Data and how Big Data analytics aids indiscovering valuable decisions. Factors for the improved quality of healthcare arediscussed which are providing patient-centric services, detecting spreading diseasesearlier, monitoring the hospital’s quality and enhancing the treatment technique.


Big Data lifecycle involves data collection, data cleaning, data classification, datamodelling and data delivery. Security is the main factor in the Big Data processing indispersed environment. The main challenges are to provide network-level security,authentication for users, nodes and application in the distributed environment,identification of the malicious hackers, etc. Author has done some work on thesecured layered data architecture for Big Data architecture. Using SSL for thecorrespondence through RPC between appropriated hubs through organize levelsecurity. Static and dynamic data can be handled by two-way communication.Attribute-based encryption method is used for transmitting the data between nodesfor the prevention of the data from the malicious users. The author in [21] discussedthe problems in the HDFS as it deals with the storage of the large files. Thefiles within the volume of 5 MB are considered as small files. The current filesystems including distributed file systems, local file systems and object-basedstorage systems all are designed for the handling of large files. For example, XFS,GPFS and HDFS all are targeting mainly the large files. In small file system, theperformance degradation will occur while processing small files through HDFS.Storage of small files will increase the internal storage space. MapReduce needsmore tasks for processing large amount of data, so it will increase overhead ofthe CPU. Author proposed that merging of small files into large file and theninputting it to HDFS will remove the problem. Storing a lot of small files intolarge files will reduce the quantity of files and increase the efficiency of retrieval ofdata. It will also decrease pressure on the disk file system. The data is increasingday by day in healthcare; its volume, velocity, variability, variety and veracity[22] are big problems these days, so it is difficult to handle large amount of datawithout a planned and efficient tool and technique. The author [23] presents a novelBig Data framework for healthcare applications. Apache Spark used a Big Dataanalytic framework, but it cannot function alone. The supporting components suchas Hadoop Core and JDK are required. The author adopts spring framework, as itis fault tolerating. Spring framework is a set of extension of the Java programminglanguage, and it is modular in nature (Fig. 3).

6 Conclusion

Big Data are referred as large volumes of data stored at the different location. Thisdata has very high velocity and complex and variable data which requires advancetools and techniques for its creation, storage and maintenance and its governance.With the increase in the volume of the data, there are many challenges for handlingof Big Data like security and privacy issues, unethical use of data, bad data quality,storage and maintenance of the data. The models and architectures used for the BigData analysation are very difficult to implement. Despite the numerous benefits ofthe Big Data, factors such as resistance to change from traditional mode to use theICT as well as security challenges are the main hindrance in effective adoption ofBig Data in healthcare.


Big DataSources

Big DataTransformation

Big DataPlatforms & Tools

Big DataAnalytics

Applications

Big DataAnalytics

▪ Hadoop

▪ MapReduce

▪ Pig

▪ Hive

▪ Jaql

▪ Zookeeper

▪ HBase

▪ Cassandra

▪ Oozie

▪ Avro

▪ Mahout

▪ Others

TransformedData

Middleware

DataWarehouse

TraditionalFormat

CSV, Tables

ExtractTransform

Load

RawData

▪ Internal

▪ External

▪ Multiple Formats

▪ Multiple Locations

▪ MultipleApplications

Queries

Reports

OLAP

DataMining

Fig. 3 Generic architecture of Big Data

References

1. Khan, R., Khan, S. U., Zaheer, R., & Khan, S. (2012). Future internet: The internet of thingsarchitecture, possible applications and key challenges, Proc. – 10th Int. Conf. Front. Inf.Technol. FIT 2012, pp. 257–260.

2. Yu, Y., Wang, J., & Zhou, G. (2010). The exploration in the education of professionals inapplied Internet of Things Engineering, ICDLE 2010–2010 4th Int. Conf. Distance Learn.Educ. Proc., pp. 74–77.

3. Da Xu, L., He, W., & Li, S. (2014). Internet of things in industries: A survey. IEEE Transactionson Industrial Informatics, 10(4), 2233–2243.

4. Gubbi, J., Buyya, R., Marusic, S., & Palaniswami, M. (Sep. 2013). Internet of things (IoT):A vision, architectural elements, and future directions. Future Generation Computer Systems,29(7), 1645–1660.

5. Ghosh, A., & Das, S. K. (2010). Coverage and connectivity issues in wireless sensor networks:A survey. Pervasive and Mobile Computing, 4(3), 303–334.

6. Wang, F., & Yuan, H. (2010). Challenges of the sensor web for disaster management.International Journal of Digital Earth, 3(3), 260–279.

7. Sookhak, M., et al. (2015). Remote data auditing in cloud computing environments: a survey,taxonomy, and open issues. ACM Computing Surveys (CSUR), 47(4), 65.

8. Abdelaziz, A., et al. (2017). Distributed controller clustering in software defined networks.PloS One, 12(4), e0174715.

9. Jia, X., Feng, Q., Fan, T., & Lei, Q. (2012). RFID technology and its applications in Internetof Things (IoT), 2012 2nd Int. Conf. Consum. Electron. Commun. Networks, pp. 1282–1285.

10. Armbrust, M., Fox, A., Griffith, R., Joseph, A., & Katz, R. H. (2010). Above the clouds: ABerkeley view of cloud computing, Univ. California, Berkeley, Tech. Rep. UCB, pp. 7–13.

11. Icu, D. L. & Icu, H. L. (2011, March). Efficient Novel Anti-collision Protocols for PassiveRFID Tags, no.

12. Wattegama, C. (2014). ICT for disaster management. Bangkok: UNDP-APDIP.


13. Chen, Z., Li, Z., Liu, Y., Li, J., & Chen, J. (2011). Quasi real-time evaluation system for seismicdisaster based on internet of things, Proc. – 2011 IEEE Int. Conf. Internet Things Cyber, Phys.Soc. Comput. iThings/CPSCom 2011, pp. 520–524.

14. Ma, Y., Liu, X., Li, X., Sun, Y., & Li, X. (2011). Rapid assessment of flood disaster loss inSind and Punjab province, Pakistan based on RS and GIS, 2011 Int. Conf. Multimed. Technol.ICMT 2011, no. Figure 1, pp. 646–649.

15. Liu, J., Wen, J., Yang, K., Shang, Z., & Zhang, H. (2011). GIS-based analysis of flood disasterrisk in LECZ of China and population exposure, Proc. – 2011 19th Int. Conf. Geoinformatics,Geoinformatics 2011, no. 40471028, pp. 0–3.

16. Seal, V., Raha, A., Maity, S., Mitra, S. K., Mukherjee, A., & Naskar, M. K. (2012). A real timemultivariate robust regression based flood prediction model using polynomial approximationfor wireless sensor network based flood forecasting systems (pp. 432–441). Berlin Heidelberg:Springer.

17. Ahmad, N., Hussain, M., Riaz, N., Subhani, F., Haider, S., Alamgir, K. S., & Shinwari, F.(2013). Flood prediction and disaster risk analysis using GIS based wireless sensor networks,a review. Journal of Basic and Applied. Scientific Research, 3(8), 632–643.

18. Sulaiman, N. A., Husain, F., Hashim, K. A., & Samad, A. M. (2012). A study on flood riskassessment for Bandar Segamat sustainability using remote sensing and GIS approach, in 2012IEEE Control and System Graduate Research Colloquium, pp. 386–391.

19. Dawod, G. M., & Koshak, N. A. (2011). Developing GIS-based unit hydrographs for floodManagement in Makkah Metropolitan Area, Saudi Arabia. Journal of Geographic InformationSystem, 03(02), 160–165.

20. Akar, Î., Kalkan, K., & Maktav, D. (2011). Determination of land use effects on flood risk byusing integration of GIS and remote sensing, Recent Adv.

21. Al-Jabari, S., Sharkh, M., & Mimi, Z. (2010). Estimation of runoff for agricultural watershedusing SCS curve number and GIS.

22. Sherief, Y. (2010). Flash floods and their effects on the development in El-Qaá plain area inSouth Sinai, Egypt, Diss. PhD dissertation, University of Mainz, Germany.

23. Fang, S., Xu, L., Zhu, Y., Liu, Y., Liu, Z., Pei, H., Yan, J., & Zhang, H. (2015). An integratedinformation system for snowmelt flood early-warning based on internet of things. InformationSystems Frontiers, 17(2), 321–335.

Internet of Things-Based Healthcare:Recent Advances and Challenges

Syed Tauhid Ullah Shah, Hekmat Yar, Izaz Khan, Muhammad Ikram,and Hussain Khan

1 Introduction

The Internet of Things (IoT) is an emerging technology consisting of a setof interconnected objects that connect anything, anyone, anyplace, anytime, anynetwork, and any service. The IoT technologies have the potential to influence theoverall business spectrum as each device and object can be recognized uniquelywithin the modern internet infrastructure, with vast benefits. These benefits normallyconsist of the advanced connectivity of systems, services, and devices that goesbeyond machine-to-machine (M2M) situations [1]. The IoT makes suitable solu-tions available for a variety of applications and services, including traffic congestion,waste management, smart cities, security, smart health, logistics, disaster services,healthcare, trade, and business control. Medical and healthcare signify one of themost striking application spaces for the IoT [2]. IoT technology has the capability toenhance medical applications such as fitness programs, elderly care, remote healthmonitoring, and management of chronic diseases. Compliance with medicationand treatment at home is another likely vital application. Consequently, differentmedical and diagnostic sensors and devices may be observed by means of objectsor smart devices, establishing an essential measure of the IoT technology. The IoThas the potential to provide enhanced user understanding and improvement in thequality of human life with a minimum cost. From the point of view of healthcaresuppliers, the IoT is capable of minimizing device downtime via remote delivery.

Over the past decade, a huge number of researchers have investigated theIoT’s abilities in terms of healthcare by considering different real-world problems.Consequently, there are currently various services and applications in the area.

S. T. U. Shah (�) · H. Yar · I. Khan · M. Ikram · H. KhanAbdul Wali Khan University, Mardan, Pakistan


153


https://doi.org/10.1007/978-3-319-96139-2_15


Leading research in the IoT for healthcare consists of new applications and services,network platforms and architecture, security, and interoperability between eachothers. Furthermore, guidelines and policies have been established for setting up IoTtechnology in the area of healthcare in various organizations and countries aroundthe world. In the future, a systematic review of the research work presently beingundertaken regarding the IoT for healthcare would be worthwhile for the numerousparties interested in further exploration of this topic. This chapter surveys the trendsin the IoT-based healthcare field and discovers concerns that need to be resolved toalter healthcare technologies using the IoT revolution. It adds to the knowledge baseunderlying this revolution by:

• Underlining different industrial endeavors to embrace IoT-based healthcareprototypes.

• Categorizing the current IoT-based healthcare network into various themes.• Providing a comprehensive discussion of IoT-based healthcare applications and

services.• Discussing fundamental technologies that can reform health technologies.• Providing a comprehensive discussion of privacy and security issues relating to

healthcare frameworks.• Discussing open research issues and challenges that need to be addressed to

create robust healthcare technologies.

2 Networks for Internet of Things (IoT)-Based Healthcare

The IoT network for healthcare (IoThNet) is an important component of theHealthcare IoT. It provides strength to the IoT, aids in the communication of healthinformation, and permits personalized communication in healthcare. Figure 1 showsthe IoThNet architecture, platform, and topology.

2.1 IoT Network for Healthcare (IoThNet) Architecture

IoThNet architecture refers to the outline for the design of the physical features,working techniques, and principals and functional organization of the IoThNet. Themain concerns relating to this have been recognized and discussed by Zhang et al.[3]. These include the interoperability of the wireless local area network and IoTgateway, secure transmission between caregivers and IoT gateways, and multimediastreaming. Several studies [4–7] have shown that 6LoWPAN (IPv6 over Low-PowerWireless Personal Area Networks) is the basic architecture for the IoThNet. In theIoThNet concept, wearables and sensors use 6LoWPAN and IPv6 (Internet Protocolversion 6) systems to communicate through the IEEE 820.15.4. After that, data aretransmitted with the help of the user datagram protocol (UDP) to sensor nodes.

Internet of Things-Based Healthcare: Recent Advances and Challenges 155

Fig. 1 The Internet-of-Things network for healthcare (IoThNet) architecture

2.2 Topology of the IoThNet

The topology of the IoThNet refers to the organization of various components ofan IoT-based network for healthcare and specifies an example of settings foundin unified medication environments. Figure 2 illustrates how a computer grid witha heterogeneous nature gathers a massive amount of sensor data and vital signs,such as body temperature, respiration rate, and blood pressure, creating an IoThNettopology. The framework harnesses the storage capability of mobile and staticelectronic devices, including smartphones, laptops, and medical terminals, creatinghybrid computing grids [8].

2.3 The IoThNet Platform

The IoThNet platform consists of both a computing platform and network platformmodel. The significance of the regulation of boundaries between stakeholdersinvolved in the design of the IoThNet is highlighted in Pang et al. [9]. A number ofstudies [3, 10, 11] have discussed concerns regarding the IoThNet platform. How-ever, these studies do not provide sufficient generalized and widespread analysisof such frameworks. A platform architecture based on semantics is presented inMiori and Russo [12]. The proposed architecture suggests semantic interoperabilitybetween various heterogeneous devices and systems.


Fig. 2 Conceptual illustration of an Internet of Things (IoT)-based pervasive healthcare solution

3 IoT-Based Healthcare Services and Applications

IoT-based health monitoring solutions are functional in a range of areas, such asobservation of prolonged diseases, care for elderly and pediatric patients, and thesupervision of private fitness and health, among others. To enhance understanding ofthis extensive topic, this chapter divides the discussion into two parts: applicationsand services. Applications are then categorized into two sets: clustered-conditionand single-condition applications. A clustered-condition application deals with alarge number of conditions or diseases, whereas a single-condition application dealswith a single condition or disease.

3.1 IoT Healthcare Services

The IoT is expected to facilitate a wide range of smart healthcare services. Theseservices individually provide a collocation of healthcare support solutions. Eachservice can be considered a platform for a wide range of applications and solutions.IoT-based protocols and services require minor adjustments to operate appropriatelyin healthcare environments. These services consist of link protocols, resource-


sharing services, notification services, and cross-connectivity and internet services.This section outlines different kinds of healthcare services based on the IoT.

The Internet of m-Health Things m-health is the combination of medical sensors,mobile and communication technologies for healthcare facilities. Istepanian etal. [13] examined implementation issues, Internet of m-health Things (m-IoT)architecture, and challenges of non-invasive glucose level sensing.

Adverse Drug Reactions An adverse drug reaction refers to harm caused by usinga medication [14]. This harm may occur after prolonged administration or a singledose of a medication or as a result of a change in its administration.

Community Healthcare Community healthcare refers to the establishment of anetwork that can cover an area such as a residential location or local community.A cooperative IoT platform-based community for rural healthcare was proposed byRohokale et al. [16].

Children’s Health Information Raising awareness and improving educationregarding children with mental and emotional problems and their needs amongstthe general public and their family members is essential [17].

3.2 IoT Healthcare Applications

In addition to IoT services, IoT healthcare applications require close attention.It must be taken into account that applications require services, but are usedby patients. Hence, applications become user-centric. This section describes thedifferent applications of healthcare based on the IoT, comprising both clustered andsingle applications (Table 1).

Glucose Level Sensing Diabetes is a metabolic disease that increases the bloodglucose level for a certain amount of time. Monitoring of glucose exposes changesin blood patterns, activities and in the formation of meal. A real-time glucose levelmonitoring scheme was introduced in Istepanian et al. [13].

Electrocardiogram Monitoring An electrocardiogram (ECG) monitors the elec-trical movement of the human heart, determining the rhythm and heart rate, QTintervals, myocardial ischemia, and diagnosis of arrhythmias. The study by Yang etal. [15] presents IoT-based solutions for ECG monitoring.

Blood Pressure Monitoring Puustjärvi and Puustjärvi [18] use the example ofblood pressure monitoring and control in developing countries as an environmentin which telemedicine would be a valuable tool.

Body Temperature Monitoring Monitoring of human body temperature is avital measure of medical services since it may be considered a vital indication ofpreservation of homeostasis [19]. Istepanian et al. [13] verified the m-IoT strategyusing body temperature sensors located in TelosB motes.


Table 1 Classification of Internet of Things (IoT)-based healthcare applications

Category Task Location Applications and devices

Robotics Monitor and helpdisabled individuals indaily life activities

Robots ADL: Meet Mr. Robin,grandma’s robot buddy [20]EADL: Social activities basedrobot [21]

Mobiledevices

Mobile health detectionand monitoring of useractivity

Carried in hands andpockets

HealthWear [22]: healthmonitoring via the skinLogbook [23]: diabetesmonitoring

Smarthomes

Centralize and automatehome tasks

Home-based systemslinked to the localhospital

Home-based healthcaresystems [24]

Wearables Monitoring and helpingdisabled individuals indaily life activities

Included in shoes,belts, and clothes

Nuubo smart shirt: smartjacket monitoring ECG,respiration, and heart rate [25]Smart watches [26]

Non-wearables

Collect user behavioralinformation at theirhomes

Included in homeobjects

Smart pillow [27]Non-contact heart ratemonitoring [28]

ADL aid to daily living, EADL electronic aids to daily living

4 IoT Healthcare Security

The IoT is growing exponentially. In the coming years, the healthcare area ispredicted to see the comprehensive adoption of IoT technology and flourish as aresult of modern eHealth IoT applications and devices. These IoT-based healthcareapplications and devices are predicted to be packed with important information,including personal healthcare information. Furthermore, these kinds of devices canbe connected to the global information network—access will be available anywhereand anytime. However, this makes IoT-based healthcare a target for hackers. Itis both essential and extremely valuable to analyze and recognize the differentfeatures of IoT privacy and security, vulnerabilities, countermeasures, and securityrequirements from the healthcare point of view to enable the IoT to adapt to thesechallenges.

4.1 Security Requirements

Security requirements in standard communications environments and IoT-basedhealthcare are similar. Hence, to achieve security facilities, an immediate focus onthis topic is required to cope with the security requirements outlined here.

Confidentiality Confidentiality ensures that medical data are not able to be accessedby malicious users or attackers.


Integrity Integrity guarantees the originality of data and that it has not been changedby an attacker during transmission.

Authentication Authentication means that the IoT device must confirm the identityof its pairing device before any communication occurs.

Availability Availability guarantees that the service will survive after a DistributedDenial of Service (DDoS) attack.

Fault Tolerance The system must be able to provide its service even in the presenceof a fault.

4.2 Security Challenges

As IoT security needs cannot be certified through traditional security schemes,innovative schemes are required to solve the novel issues encountered in the IoT.Various challenges for the IoT-based healthcare services are discussed here.

Computational Limitations Normally, IoT-based healthcare devices contain low-speed processing units. The core of these devices (the central processing unit [CPU])is not very influential in terms of performance and speed and does not perform theexpensive computational operations.

Memory Limitations The majority of devices are equipped with a lower amountof built-in memory and can be activated with an embedded operating system (OS).

Energy and Mobility Limitations IoT-based healthcare devices are dynamic andequipped with small health devices and batteries. As various networks have differentconfigurations and settings, a mobility–complement security algorithm is required.

5 IoT Healthcare Technologies

There are various supporting technologies for IoT medication frameworks and it ischallenging to a clear scope. In this section, key technologies with the ability totransform IoT-based healthcare are illustrated.

5.1 Wearables

Health enhancements and patient engagement can be assisted by using wearablehealthcare devices. Using such technology can result in three major advantages:gamification, target-oriented healthcare, and connected information.


5.2 Cloud Computing

The incorporation of HealthcareIoT with cloud computing should offer serviceswith access to mutual resources, allowing different needs to be met via connectionsthrough the cloud network.

5.3 Networks

Different networks, including long-range communications (cellular networks) andshort-range communications (WLANs [wireless local area networks], WBANs[wireless body area networks], 6LoWPANs, WSNs [wireless sensor networks], andWPANs [wireless personal area networks]), are a vital requirement of the physicalinfrastructure of HealthcareIoT.

5.4 Big Data

Big data contains large volumes of vital health data created by different medicaldevices and allows the creation of tools to enhance the proficiency of healthmonitoring, staging, and diagnosis.

6 IoT Healthcare: Current Issues and Challenges

Various researchers have focused on implementing and designing different IoThealthcare frameworks and solving numerous architectural complications relatedto those frameworks. However, there are still numerous open research issues andchallenges that need to be properly addressed. This section outlines some of thesechallenges.

Cost Analysis Researchers need to consider the creation of low-cost prototypes ofIoT-enabled medication solutions; however, to date, no study has considered thistopic.

Continuous Monitoring In some conditions, long-term patient health monitoring isrequired. To achieve this, continuous logging and monitoring is essential.

Identification Healthcare organizations and hospitals often deal with a huge numberof patients, with multiple support staff performing various duties. To achieve properdata management in these situations, accurate identification of patients and staff isessential.


Mobility The IoT healthcare solutions must be capable of assisting the mobility ofpatients because they are linked anytime, anywhere. The mobility specifications areused to connect different patients to different networks.

7 Conclusion

Researchers around the globe have started to discover numerous technologicalsolutions that will improve healthcare in a way that enhances current servicesby bringing together the capabilities of the IoT. This chapter investigates variousfeatures of the IoT healthcare framework and discusses different healthcare networkarchitectures that support access to this IoT framework and aid the reception andtransmission of medical data. Furthermore, the chapter notes research that has beenundertaken regarding how the IoT can assist in elderly and pediatric care, privatehealth, fitness management, and chronic disease supervision. To aid understandingof privacy and security issues relating to IoT healthcare, this chapter outlinesmany security requirements and issues and notes various research problems. Inaddition, IoT and eHealth regulations and policies relating to various stakeholdersthat are significant in accessing IoT-based healthcare technologies are presented.To conclude, the outcomes of this chapter are predicted to be beneficial for healthprofessionals, researchers, engineers, and policymakers operational in the field ofhealthcare technologies and the IoT.

References

1. Höller, J., Tsiatsis, V., Mulligan, C., Karnouskos, S., Avesand, S., & Boyle, D. (2014). Frommachine-to-machine to the internet of things: Introduction to a new age of intelligence.Amsterdam: Elsevier.

2. Pang, Z. (2013). Technologies and architectures of the internet-of-things (IoT) for health andwell-being. KTH Royal Institute of Technology: Diss.

3. Zhang, X. M., & Zhang, N. (2011). An open, secure and flexible platform based on internetof things and cloud computing for ambient aiding living and telemedicine. 2011 internationalconference on computer and management (CAMAN). IEEE; 2011.

4. Imadali, S., et al. (2012). eHealth service support in IPv6 vehicular networks. IEEE 8thinternational conference on wireless and mobile computing, networking and communications(WiMob). IEEE; 2012.

5. Jara, A. J., Zamora, M. A., & Skarmeta, A. F. (2012). Knowledge acquisition and managementarchitecture for mobile and personal health environments based on the internet of things. IEEE11th international conference on trust, security and privacy in computing and communications(TrustCom). IEEE; 2012.

6. Doukas, C., & Maglogiannis, I. (2012). Bringing IoT and cloud computing towards pervasivehealthcare. Sixth international conference on innovative mobile and internet services inubiquitous computing (IMIS). IEEE; 2012.

7. Jara, A. J., et al. (2010). Drugs interaction checker based on IoT. Internet of Things (IOT).IEEE; 2010.


8. Viswanathan, H., Lee, E. K., & Pompili, D. (2012). Mobile grid computing for data-andpatient-centric ubiquitous healthcare. First IEEE workshop on enabling technologies forSmartphone and Internet of Things (ETSIoT). IEEE; 2012.

9. Pang, Z., et al. (2013). Ecosystem analysis in the design of open platform-based in-homehealthcare terminals towards the internet-of-things. 15th international conference on advancedcommunication technology (ICACT). IEEE; 2013.

10. Jara, A. J., Zamora-Izquierdo, M. A., & Skarmeta, A. F. (2013). Interconnection frameworkfor mHealth and remote monitoring based on the internet of things. IEEE Journal on SelectedAreas in Communications, 31(9), 47–65.

11. Vazquez-Briseno, M., et al. (2012). A proposal for using the internet of things concept toincrease children’s health awareness. 22nd international conference on electrical communica-tions and computers (CONIELECOMP). IEEE; 2012.

12. Miori, V., & Russo, D. (2012). Anticipating health hazards through an ontology-based,IoT domotic environment. Sixth international conference on innovative mobile and internetservices in ubiquitous computing (IMIS). IEEE; 2012.

13. Istepanian, R. S., et al. 2011. The potential of Internet of m-health Things “m-IoT” for non-invasive glucose level sensing. Engineering in Medicine and Biology Society, EMBC, 2011annual international conference of the IEEE. IEEE.

14. ICH Expert Working Group. (1996, April). Guidance for industry-E6 good clinical practice:consolidated guidance. Silver Springs (MD): U.S. Department of Health Human Services, USFood and Drug Administration.

15. Yang, G., et al. (2014). A health-iot platform based on the integration of intelligent packaging,unobtrusive bio-sensor, and intelligent medicine box. IEEE Transactions on Industrial Infor-matics, 10(4), 2180–2191.

16. Rohokale, V. M., Prasad, N. R, & Prasad, R. (2011). A cooperative Internet of Things(IoT) for rural healthcare monitoring and control. 2nd international conference on wirelesscommunication, vehicular technology, information theory and aerospace & electronic systemstechnology (Wireless VITAE). IEEE; 2011.

17. Awareness Day 2014 Activities by Program Type. http://www.samhsa.gov/sites/default/files/children-awareness-dayactivities-by-program-2014.pdf. Accessed 7 Dec 2014.

18. Puustjärvi, J., & Puustjärvi, L. (2011). Automating remote monitoring and information therapy:An opportunity to practice telemedicine in developing countries. IST-Africa conferenceproceedings, 2011. IEEE.

19. Ruiz, M. N., García, J. M., & Fernández, B. M. (2009). Body temperature and its importanceas a vital constant. Revista de enfermeria (Barcelona, Spain), 32(9), 44–52.

20. Monks, Keiron. Meet Mr. Robin, grandma’s robot buddy. CNN. http://www.cnn.com/2014/05/16/tech/innovation/meet-mr-robin-grandmas-robot/

21. Brose, S. W., et al. (2010). The role of assistive robotics in the lives of persons with disability.American Journal of Physical Medicine & Rehabilitation, 89(6), 509–521.

22. Leonhardt, S. (2006). Personal health care devices. In S. Mukherjee et al. (Eds.), Amlwarehardware technology drivers of ambient intelligence (pp. 349–370). Dordrecht: Springer.

23. http://www.ihealthlabs.com/glucometer/wireless-smart-gluco-monitoring-system/24. http//beclose.com/press-070912.aspx25. http://nuubo.es/?q=en/node/16226. http://www.idownloadblog.com/2015/05/25/applewatch-blood-oxygen-saturation-hardware/27. https://itunes.apple.com/us/app/pillow-smart-sleep-cycle-alarm/id878691772?mt=828. Kagawa, M., Yoshida, Y., Kubota, M., Kurita, A., & Matsui, T. (2012). Non-contact heart

ratemonitoring method for elderly people In bed with random body motions using 24 GHzdual radars located beneath the mattress in clinical settings. Journal of Medical Engineering &Technology, 36, 344–350.

http://www.samhsa.gov/sites/default/files/children-awareness-dayactivities-by-program-2014.pdf

http://www.cnn.com/2014/05/16/tech/innovation/meet-mr-robin-grandmas-robot/

http://www.ihealthlabs.com/glucometer/wireless-smart-gluco-monitoring-system/

https://www.alarm.com/productservices/wellness.aspx/press-070912.aspx

http://nuubo.es/?q=en/node/162

http://www.idownloadblog.com/2015/05/25/applewatch-blood-oxygen-saturation-hardware/

https://itunes.apple.com/us/app/pillow-smart-sleep-cycle-alarm/id878691772?mt=8

Index

AAging

challenges, 99mobility and extensive care, 99–100nuclear cataract, 36

Ambient assisted living (AAL), 99–100Augmented reality

clinical administration and prescriptionlayer, 105

image capturing and object recognitiontechniques, 102–103

patient drug compliance layer, 105–106pharmacy vendor management, 104

Smart Assist architecture, 104, 105Augmented virtuality, mobile apps, 69–71Automated blood analyzer, 88Automatic detection and grading of cataract

digital images, 40literature survey analysis, 43–44LOCS-III/Wisconsin grading protocol, 42retinal image, 39–40retro-illumination images, 38–39slit-lamp images, 37–38ultrasonic nakagami images, 41

BBig Data

architecture, 149–151attribute-based encryption method, 150challenges and limitations

data governance, 148ethical and moral challenges, 149

security and privacy issues, 149security threats, 148

characteristics, 144–145components, 144concept, 146dimensions of, 143healthcare cloud, 148Health-CPS, 144IoT, 162lifecycle, 150MapReduce, 150organizational challenges, 143tools and techniques for analysing, 146,

147Blood

cancer cells prediction, 89CBC, 88count, 87elements, 87–88functions, 87image contouring technique, 89image segmentation techniques, 89tests, 88

Blood cell counting and segmentationCIELAB, 89convex hull and convex deficiency, 93experimental results, 94extraction of the white blood cells, 90–91future perspectives, 95–97image segmentation techniques, 89intensity thresholding, 89K-means algorithm, 91–92LoG edge detection, 89performance evaluation measures, 93–94

© Springer Nature Switzerland AG 2019F. Khan et al. (eds.), Applications of Intelligent Technologies in Healthcare,EAI/Springer Innovations in Communication and Computing,https://doi.org/10.1007/978-3-319-96139-2

163

https://doi.org/10.1007/978-3-319-96139-2

164 Index

Blood pressure (BP) ranges, 1Body sensor networks (BSN), 49

CCardiovascular disease (CVD)

blood vessels/heart, 2detection techniques, 117diagnosis methods, 2heart rate, 1heart sound analyzer, 118hybrid classifier

KNN, 124Naïve Bayes, 124–126SVM, 124

mortality rate, 117PCG signal, 2, 3

Cataractautomatic detection and grading

digital camera images, 40retinal image analysis, 39–40retro-illumination images, 38–39slit-lamp images, 37–38ultrasonic B-scan and Nakagami

imaging techniques, 41diagnosis, 36grading protocols, 37vs. normal vision, 36risk factors, 35types, 36

Chest radiographs segmentationactive shape model, 26dataset, 30Dice coefficient, 30, 31JSRT dataset, 26, 31kernel mapping, 26lung region segmentation, 26proposed methodology

flowchart, 27graph cuts iteration, 30MAP formulation, 27, 28Mercer’s theorem, 28performance evaluation, 32update of labels, 30update of region parameters, 29–30visualization, 31

segmented mask vs ground truth, 31CIE L*a*b* (CIELAB), 89Classification and regression trees (CART)

algorithm, 14Cloud-assisted, IoT healthcare

patient respiration monitoring, 80proposed schemes, 81

Cloud computing, 144

Complete blood count (CBC), 88shortcomings, 88test methods, 88

Convex hull and convex deficiency, 93Cortical cataract (CC), 36

DData governance, 148Defender cells, 88Deoxyribonucleic acid (DNA) sequence

bioinformatics and computational biology,56

dataset, 61experimental environment, 61Parallel vector space model

accuracy analysis, 62, 63cosine values, 59–60efficiency analysis, 62finding terms frequency, 58–59speed analysis, 63

Vector space model, 57Dimensionality reduction, 18–19

EElectrocardiogram (ECG), 117Electronic health records (EHRs), 49Erythrocytes. See Red blood cellsEye survey

automatic detection and grading of cataractdigital images, 40literature survey analysis, 43–44LOCS-III/Wisconsin grading protocol,

42retinal image, 39–40retro-illumination images, 38–39slit-lamp images, 37–38ultrasonic Nakagami images, 41

diagnosis of cataract, 36–37formation of cataract, 36normal vs. cataract vision, 36screening methods, 35visual acuity tests, 35

FFourier domain OCT (FD OCT) systems,

113–114

HHadoop, 149Health monitoring scheme

feature extraction, 83–84

Index 165

MIMIC II database, 84proposed approach, 84signal enhancement, 81–82watermarking, 82workload transmission, 85

Heart attacks, 117Heart beat analysis

artificial neural network, 19classification, 19dataset, 15–16dimensionality reduction, 18–19feature extraction

energy-based features, 18FFT and spectral features, 16–17probability-based features, 17–18

low-pass filtering and decimation, 14nonnegative matrix factorization, 15normal/abnormal classification, 14PCG signal, 14principal component analysis

classification rates with and without, 20energy-based features, 21individual and combined features, 22,

23probability-based features, 21spectrum-based features, 22, 23

Springer segmentation algorithm, 14Heart sound analyzer, 118Heart sound identification system

block diagram, 120PCG dataset description, 119performance, 125preprocessing, 119–121results, 126segmentation, 121–123training phase, 125

Hemoglobin, 87Hidden semi-Markov models (HSMM), 4,

14Hybrid classifier, CVD

KNN, 124Naïve Bayes, 124–126SVM, 124

IImage segmentation techniques, 89Internet of Things (IoT)

accountability, 52ambient assisted living, 50attack and threat agent, 52benefits, 155body sensor networks, 49

cloud-assistedpatient respiration monitoring, 80proposed schemes, 81

conceptual design, 78cyberbullying purposes, 52DDOS attacks, 51devices leaking private information, 51electronic health records, 49enhance medical applications, 155guidelines and policies, 156healthcare applications, 159–160healthcare services, 158–159IoThNet

architecture, 156, 157platform, 157topology, 157

literature review, 79remote monitoring system, 49research issues and challenges, 162–163robust privacy-preserving techniques, 53security

challenges, 161hazards, 51requirements, 160–161

sustainable environments, 48technologies

Big Data, 162cloud computing, 162networks, 162wearable healthcare devices, 161

traditional hiding techniques, 51traditional privacy principles, 51vulnerabilities, 51

IoT network for healthcare (IoThNet)architecture, 156, 157platform, 157topology, 157

KK-means algorithm, 92

LLaplace of Gaussian (LoG) edge detection, 89Leukocytes. See White blood cellsLOCS-III/Wisconsin grading protocol, 37, 39,

42Logistic regression classification model, 119

MMapReduce, 150MedHelp, 102

166 Index

MedLink, 102Mel frequency cepstral coefficients (MFCCs),

14Mobile fitness applications

augmented virtuality, 69–71features, 68, 73systematic review, 68–70virtual reality, 71–72

MyMed, 102MyMedSchedule, 101

NNonnegative matrix factorization (NMF), 15Nuclear cataract (NC), 36

OOpenCV library, 104Optical coherence tomography (OCT)

disease detection, 109Fourier domain OCT systems, 113future perspectives, 114modalities, 112robustness and efficiency, 109in segments and slices, 109terminologies, 110–111time domain OCT systems, 112

PParallel vector space model (PVSM)

accuracy analysis, 62, 63cosine values, 59–60efficiency analysis, 62finding terms frequency, 58–59speedup analysis, 63splitting query and documents, 58weight calculation, 59

Phonocardiogram (PCG) signals, 117–118with annotations, 3cardiac cycle, 3CIC filtering, 4classification, 8–9dataset, 5, 119feature extraction, 7–8heart beat analysis, 14KNN classifier, 9localization, 7machine learning algorithms, 4methodology, 5–6normal and abnormal, 123, 127

Pascal and PhysioNet challenge, 5proposed methodology, 6

quality assessment, 6–7Savitzky-Golay filter, 119–121segmentation, 121–122time-domain filtering, 4wavelet coefficients, 121, 122

Photoplethysmogram (PPG), 117Plasma, 87Platelets, 88Posterior subcapsular cataract (PSC), 36Principal component analysis (PCA)

classification rates with and without, 20energy-based features, 21individual and combined features, 22, 23probability-based features, 21spectrum-based features, 22, 23

Probability density function (PDF), 17

RRed blood cells, 87

convex hull and convex deficiency, 93Delaunay triangulations, 89image binarization, 91–92morphological methods, 89

Red Blood Cell Segmentation Using Maskingand Watershed Algorithm, 89,93–94

RxmindMe, 102

SSavitzky-Golay filter, 120Smart Assist, 102

architecture, 104, 105functions, 104

Smartphone-based drug complianceaugmented reality

clinical administration and prescriptionlayer, 105

image capturing and object recognitiontechniques, 102–103

patient drug compliance layer, 105–106pharmacy vendor management, 104Smart Assist architecture, 104, 105

clinical outcomes, 100mobile apps, 100SMS and electronic devices, 100–101

Snake algorithm, 89Spectral domain OCT (SD OCT) systems, 113Spring framework, 150SQLite database, 104Support vector machine (SVM), 14Swept source OCT (SD OCT) systems,

113–114

Index 167

TThrombocytes. See PlateletsTime domain OCT (TD OCT) systems, 112

VVaccination

booster shots, 133control flow, 136, 137directly transmitted virus, survival of

extensions, 132factors affecting, 132

efficacy period, 136experimental design, 137, 138FP Vacc, 134, 138–139future perspectives, 140herd immunity, 133N Vacc, 134, 138–139

open and closed population, 133parameter sweeping, 137periodic vaccination, 133RP Vacc, 134, 138–139

Vaccine-preventable death, 131Vaccine-preventable diseases, 131Vector space model (VSM), 57Virtual reality, mobile apps, 71–72

WWearable healthcare devices, 161White blood cells, 88

extraction, 90–91intensity thresholding, 89removal process, 89Snake algorithm, 89

Wold’s decomposition model, 89