[ieee 2011 ieee international conference on control system, computing and engineering (iccsce) -...

Discrimination of Vision Impairments Using Single Trial VEPs

Vikneswaran Vijean, M.Hariharan, Sazali Yaacob School of Mechatronic Engineering

University Malaysia Perlis Perlis, Malaysia

[email protected]

Abstract— Analysis of Visually evoked potential (VEP) in the investigation of ocular diseases is gaining interests from researchers all over the world. VEP is an electrical signal generated by the brain (Occipital Cortex) in response to a visual stimulus. By analyzing these responses, the abnormalities in the visual pathways in a person can be detected. Traditionally, the amplitude and the latency values were considered for the analysis. This study is intended to investigate the frequency domain based features of single trial VEPs in discriminating between subjects with normal vision from those having vision impairments. Four different classifiers, Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), k Nearest Neighbor (kNN) and the Support Vector Machine (SVM) are used for the investigation. The proposed method shows promising results for the discrimination of vision impairments.

Keywords-visually evoked potential; vision impairement; frequency domain feature .

I. INTRODUCTION Visually evoked potentials have been widely used as a

reliable indicator for the diagnostics of ocular diseases. Although measured from the scalp, the VEP differs from the EEG signal in several ways [1]. First of all, the VEP corresponds to a specific response of occipital lobe, part of the brain which is involved in receiving and interpreting the visual signals, while the EEG signifies the ongoing activity of wide areas of the cortex [1,2]. The VEP is also more sensitive towards the changes in the stimulus conditions compared to the EEG. The commonly used visual stimuli is a black and white checkerboard pattern, and is presented to the patients in order to study the responses of VEP although other types of stimulus can also be used such as flash and line gratings depending on the type of diseases to be investigated. Doctors usually refer their patients to perform the VEP test if the patients are experiencing changes in vision that can be due to problems along the visual nerves pathways [2]. Analysis of the VEP would help to determine the nature of the diseases. Although VEP is particularly useful in detecting past optic neuritis, it can also be used to detect other ocular diseases such as glaucoma, multiple sclerosis, ocular hypertension, macular degeneration, colour blindness, Parkinson’s disease, idiopathic intracranial hypertension and cataract [1-10]. The commonly extracted parameters for analysis from the VEP are the amplitude and latency values. The amplitudes examined are the N75, P100, and N125 where ‘N’ denotes the negative peak, ‘P’ for positive

peak and the numbers (75,100, 125) indicating the time in microseconds. The latency value is measured as the time taken for a visual stimulus to travel from the eye to the occipital cortex [2,4,11-13]. Traditionally, ensemble averaging technique is used to separate the VEP from the background noise to perform the analysis. In ensemble averaging, a minimum of 65 single trials are averaged to obtain the VEP response. This however, results in the information’s related with variations between the single trials to be lost [14]. Hence, in current research work, the signal processing techniques are applied to the single trial VEPs in an effort to develop an alternative diagnostic method for distinguishing the vision impairments. Three control subjects with perfect vision and three subjects with vision impairments are used for the analysis. The subjects with vision impairments in this paper refer to those who are near sighted and need visual aids to see clearly.

II. VEP RECORDINGS PROCEDURES The experiment is conducted in accordance to the standard

procedure established by the international society for clinical electrophysiology of vision (ISCEV). For the analysis, VEP data is collected from 12 eyes of 6 subjects aged between 24 to 30 years (3 control subjects and 3 subjects with vision impairment and having to wear visual aids). Since the VEP originates from the occipital cortex, each “eye” is treated as an individual “case”. The data is taken from subjects that are seated 60cm in front of a black and white checkerboard pattern that is displayed on a HP L1908w LCD monitor. Two different check sizes, 1 degree and 0.25 degree are used in this study. Subjects are asked to focus their gaze on the red fixation at the centre of the screen while the checkerboard changes phase abruptly at a rate of 2 reversals per second. The monocular VEP recordings are done by placing cup-shaped gold plated electrodes on the scalp overlaying the occipital region. A light tight opaque patch is used for this purpose. The interelectrode resistance was kept below 3kΩ. The active electrodes are placed at Oz and O1 position, while the reference electrode is placed at Fz and the ground electrode is placed on the vertex (Cz position) according to the international 10-20 system. The signals are band pass filtered in the range of 0.3 Hz to 100 Hz. The three subjects with vision impairments are asked to wear the visual aids during the recordings.

2011 IEEE International Conference on Control System, Computing and Engineering

978-1-4577-1642-3/11/$26.00 ©2011 IEEE 182

The experimental data is collected in terms of blocks of trials. One block of trial is the continuous collection of 40 trials displayed one after other. In a typical experiment, 3-4 blocks of trials are recorded. In each block of trials, the eye blink trials are eliminated. Samples are recorded at 1 KHz, for a period of 20 seconds. Sixty five artifact free trials of data per eye of the subject are used in this analysis.

III. PREPROCESSING AND FEATURE EXTRACTION

Since the VEP is smaller in amplitude, it is more likely to be corrupted by noise. The raw signal must be preprosesed to remove the unwanted noise from the signal. The most common artifact present in the VEP signals are the eye blinks. This artifact is removed by simple thresholding method where the trials that are found to have amplitudes more than 100 v is discarded from the experiments using the written matlab program. The treshold value of 100 v is used because the blinking will produce 100-200mV potentials [15]. The single trial VEPs are then baseline calibrated by removing the mean from the samples.

The preprossesed signals are decomposed into five different frequency bands, delta (0.5Hz-4Hz), theta (4Hz-8Hz), alpha (8Hz-16Hz), beta (16Hz-32Hz) and gamma (32Hz-64Hz) using the elliptic digital filter. The filter order is chosen as to give a maximum of 0.1db passband ripple and a minimum of 20db stopband attenuation at ±0.5Hz in respective frequency bands[16]. The ‘filtfilt’ funtion is used to perform the filtering twice, forward and reverse filtering as to cancel the phase non-linearity of the elliptic filtering. The signals are converted to frequency domain using the fast fourier transform algorithm. Hanning window is used to reduce the spectral leakages. The statistical frequency domain features are then extracted from the trials for classification. The features used in the analysis are the spectral power, spectral energy, entropy [17], standard deviation (Std) and the kurtosis of the single trial VEP’s and are shown in equation 1-5. The features are computed using a 0.5 second window.

∑ ==

n

i ixnwerSpectralPo

1

21 (1)

∑ ==

n

i ixergySpectralEn1

2 (2)

)(log)( 101 in

i i xpxpEntropy ∑ =−= (3)

21

)(1 xxn

Std in

i−= ∑ =

(4)

44 )/()( StdxxKurtosis ∑ −= (5)

where x is the samples, x is the mean of x , n is the number of samples in a 0.5 second window, )( ixp is a probability of random phenomena ix and is approximated by the difference

of spectral components and mean value, ][][)( isisxp i −≅

The extracted features are trained using the Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), k-nearest neighbor (kNN) and Support Vector Machine (SVM) classifiers. Ten-fold cross-validation technique is performed to compare the accuracy of the classifiers. The whole dataset is randomly divided into 10 mutually exclusive and approximately equal size subsets. The classifiers trains and tests the data for 10 times. In each case, a subset is used as testing data, while the remaining subsets form the training data.

IV. CLASSIFICATION

A. Linear Discriminant Analysis LDA is a well known-scheme for feature classification and

is used in this study for 2-class problem. The LDA classifier reduces the dimensional space while preserving much of the class discriminatory information [18]. This is done by projecting the data onto a lower-dimensional vector space in such a way that the ratio of the inter-class distance and the intra-class distance is maximized, thus achieving maximum discrimination between-class. The maximum likelihood method is used to fit the parameters of this classifier to the available data. The objective of this is to find a linear transformation that gives maximum class separability. The linear transformation or the discriminant function is computed as [19,20]:

)ln(21 11

iTiwi

tkwii PSxSf +−= −− μμμ (6)

where

iμ is the mean feature in group i (i=1 and 2).

kx , x is the feature of all data. k represents one feature.

iP is the total sample of each group divided by total samples.

Originally, the mean of class1 ( 1μ ), class2 ( 2μ ) and mean, 3μ of the total data set were computed, and is shown in

equation 7. 1P and 2P were prior probabilities of class 1 and class 2 respectively.

22113 μμμ ×+×= PP (7)

Class discrimination in LDA is measured by using within-class scatter wS and between-class scatter BS . The wS is calculated using equation 8.

2211 covcov ×+×= PPSw (8)

Hence, covariance of both the classes is symmetrical. Covariance matrix is calculated using equation 9 while the BS is calculated using Eq.10. can be considered as the covariance of data set by replacing mean vectors of each class with mean vectors of total data set.

Tiiiii xx ))((cov μμ −−= (9)

∑ −×−=i

TiiBS )()( 33 μμμμ (10)


183

All feature data are transformed into discriminant function. Training data and the prediction data are drawn into new coordinate. Difference between if is used to determine the Class label of prediction data.

B. Quadratic Discriminant Analysis A more generalized version of LDA is called the QDA, provided that there are only two classes of points and the measurements are normally distributed. However, in QDA the assumption that the covariance of each class is identical is not taken into consideration. Further, the surface that separates the subspaces will be a conic section ( parabola, hyperbola, etc.) [21].

C. k-nearest neighbor kNN is a simple, supervised algorithm that employs lazy

learning [19,20]. It classifies the test samples based on majority of k-nearest neighbor category. The kNN category is determined by finding the minimum distance between the test samples and each of the training sets. Each of the query instances (test VEP signal) is compared with each of the training instance (training VEP signals). The Euclidean Distance measure is used to find the closest members of the training set to the test class being examined. The label of a class is determined from the kNN category using majority voting. Equation 11 shows the calculation of Euclidean Distance.

∑=

−=N

iiiE babaD

1

2)(),( (11)

where a and b are the training and testing VEP signals composed of N features. The effect of different neighborhood in the classification results is studied by varying the k values from 1 to 4.

D. Support Vector Machine The SVM classifier operates by finding a separating hyper plane between two classes, such that the minimal distance with respect to the training vectors is maximum. The non-linear SVM is implemented by applying kernel trick to maximum-margin hyper planes. The feature space is transformed into higher dimensional space where the maximum-margin hyper plane is found. The following explains about the SVM theory [22]. For the training data set of

niyRxyxyxyx in

inn ,...,2,1,1,1,),,)...(,(),,( 2211 =+−∈∈

the training vector membership is;

⎩⎨⎧

−=−≤++=+≥+

1,1).(1,1).(

i

iybxWybxW

The distance between the hyper planes denoted by md is expressed as W/2=dm

Adopting the maximum margin criterion, the primal optimization problem is equivalent to:

.,...,2,1,1)).((,W21 2

, nibxWyMin iibW =+≥+ (12)

Applying penalty parameters to allow misclassification, the optimization problem (12) becomes

∑ =+

n

i ibW CMin1

2,, ,W

21 εε

nibxWy iiii ,...,2,1,0,1)).(( =≥−≥+ εε (13)

Solving the problem (13) using Langrangian dual problem method, the decision function becomes

∑ =+=

n

i iii bxxyxf1

)).(sgn()( α (14)

Extending the linear SVM using kernel function results in non-linear SVM as

∑ =+=

n

i iii bxxkyxf1

)).(sgn()( α (15)

where ))().(().(( xxxxk ii ϕϕ=

The RBF kernel used is expressed as 22 /),(( σxxexxk ′−−=′

where the σ value is set to 2, after searching through the 10-fold cross validation method for optimum parameter settings. All the feature extraction and classification algorithms are developed under MATLAB 7.0 environment.

E. k-fold cross validation The k-fold cross validation method is used to evaluate and

compare the LDA, QDA, kNN and SVM algorithms. Using this method, the data are divided into two segments; one for training and one for testing the classification algorithm. The data are at first divided into k equally sized segments. The k iterations of training and testing are performed in such a way that for every iteration a different segment of data is held-out for validation. The remaining k-1 segments are used for training. This method is used to avoid the overlap between the training and validation sets as they tend to lead to over estimation of the performance of the classifiers. Hence, the results obtained using k-fold cross validation method can provide accurate performance estimation of each algorithm. The value chosen for the analysis is k=10, as it utilizes 90% of the data for predictions and is more likely to be generalizable for the full data [23].


184

V. RESULTS AND DISCUSSIONS The features are binary normalized prior to training and

testing. The results obtained from the classification for LDA, QDA and kNN classifiers are presented in table 1 and the best results for each feature is highlighted. The results for individual check sizes are discussed below.

A. 0.25 degree checks From the table1, it can be seen that the standard deviation

derived from the single trial VEP’s gives the best results for all the classifiers. Features such as the power, energy and the entropy are poorly classified compared to the Std and kurtosis. The kNN algorithm with k values of 1 and 2 is able to correctly classify majority of the testing instances for Std and kurtosis features.

B. 1 degree checks The Std feature is able to produce 100% accurate

classification through kNN classifier while the SVM classifier with RBF kernel produced 99.87% accuracy. The repeatability of the results for kNN classifier shows the reliability of the Std feature derived from the frequency domain in classification of vision impairment.

The classification results for the 1 degree checks are slightly better than the 0.25 degree checks. This is in accordance to the findings by [24] that the larger checks produce much clear VEP recordings. Although LDA and QDA classifiers performance is low for majority of the features, the kNN algorithm was able to produce more than 90% accuracy for both the checks. This implies that the non linear classifiers are more suited for the discrimination of vision impairments than the linear classifiers.

VI. CONCLUSION The performance of the frequency domain based features

in the classification of vision impairment using single trial VEP is evaluated. The features extracted are tested using LDA, QDA, kNN and SVM algorithms. The results obtained shows that larger check sizes are more useful in obtaining reliable VEP responses. Perfect classification of 100% achieved by kNN classifier for the Std feature demonstrates the usefulness of the statistical features in the discrimination of vision impairments. The proposed method shows promising results in the investigation of vision impairments and can be further extended to the diagnostics of ocular diseases. In future, the time-frequency domain based analysis can also be carried out for the single trial VEP in the effort to develop an automated diagnostics system for ocular diseases.

ACKNOWLEDGMENT

The authors thankfully acknowledge the short-term research grant (No. 9001-00192) from Universiti Malaysia Perlis(UniMAP), Perlis, Malaysia.

REFERENCES [1] S. Sokol, "Visually evoked potentials: theory, techniques and clinical

applications," Survey of ophthalmology, vol. 21, ,1976, pp. 18-44. [2] Visual Evoked Potential (VEP) Summary

http://www.virtualmedicalcentre.com/healthinvest [3] S. L. Graham, A. I. Klistorner, and I. Goldberg, "Clinical application

of objective perimetry using multifocal visual evoked potentials in glaucoma practice," Archives of ophthalmology, vol. 123, 2005, pp. 729-739.

[4] G. Holder, "Electrophysiological assessment of optic nerve disease," Eye, vol. 18,, 2004, pp. 1133-1143.

[5] J. V. Odom, R. Hobson, J. T. Coldren, G. M. Chao, and G. W. Weinstein, "10-Hz flash visual evoked potentials predict post-cataract extraction visual acuity," Documenta ophthalmologica, vol. 66, 1987, pp. 291-299.

[6] Y. H. Shih, Z. J. Huang, and C. E. Chang, "Color pattern-reversal visual evoked potential in eyes with ocular hypertension and primary open-angle glaucoma," Documenta ophthalmologica, vol. 77, 1991, pp. 193-200.

[7] S. Bass, J. Sherman, I. Bodis-Wollner, and S. Nath, "Visual evoked potentials in macular disease," Investigative ophthalmology & visual science, vol. 26, 1985, p. 1071.

[8] A. Kesler, V. Vakhapova, A. D. Korczyn, and V. E. Drory, "Visual evoked potentials in idiopathic intracranial hypertension," Clinical neurology and neurosurgery, vol. 111, 2009, pp. 433-436.

Check Size Classifier\Features Power Energy Entropy std kurtosis

0.25 deg

LDA 76.33 76.42 77.00 87.88 84.92QDA 77.29 77.12 79.59 98.83 87.64

kNN(1) 88.59 88.97 89.23 99.87 98.85 kNN(2) 88.85 88.72 90.38 99.87 98.85 kNN(3) 87.82 87.31 87.95 99.62 97.95 kNN(4) 88.21 88.08 90.26 99.62 98.21 SVM 87.05 86.79 88.21 99.74 97.56

1 deg

LDA 78.77 78.60 79.08 92.54 82.56QDA 81.36 81.14 83.01 99.04 86.74

kNN(1) 89.62 88.85 87.95 100.00 99.10 kNN(2) 89.10 89.74 88.33 100.00 98.97 kNN(3) 90.26 90.77 89.49 100.00 97.44 kNN(4) 90.51 90.26 89.62 100.00 97.44 SVM 87.56 87.56 89.36 99.87 97.44

TABLE 1: CLASSIFICATIONS RESULTS


185

[9] D. Regan and D. Neima, "Visual fatigue and visual evoked potentials in multiple sclerosis, glaucoma, ocular hypertension and Parkinson's disease," Journal of Neurology, Neurosurgery & Psychiatry, vol. 47, 1984, p. 673.

[10] S. Jones, "Visual evoked potentials after optic neuritis," Journal of neurology, vol. 240, 1993, pp. 489-494.

[11] R.Sivakumar , G. R. "Identification of intermediate latencies in transient visual evoked potentials”, Academic Open Internet Journal, Volume 17, 2006.

[12] J. V. Odom, M. Bach, M. Brigell, G. E. Holder, D. L. McCulloch, and A. P. Tormene, "ISCEV standard for clinical visual evoked potentials (2009 update)," Documenta ophthalmologica, vol. 120, pp. 111-119.

[13] R. Diem, A. Tschirne, and M. Bahr, "Decreased amplitudes in multiple sclerosis patients with normal visual acuity: a VEP study," Journal of clinical neuroscience, vol. 10, pp. 67-70, 2003.

[14] R. Quian Quiroga, "Obtaining single stimulus evoked potentials with wavelet denoising," Physica D: Nonlinear Phenomena, vol. 145, pp. 278-292, 2000.

[15] Palaniappan, R., P. Raveendran, and S. Nishida," Multi-Channel Noise Reduced Visual Evoked Potential Analysis", C ( ) Vol. 123, No. 10, 2003, pp. 1721-1727.

[16] R. Palaniappan, "Vision Related Brain Activity for Biometric Authentication," pp. 3227-3231.

[17] K. Ekštein and T. Pavelka, "Entropy And Entropy-based Features In

Signal Processing," 2004. [18] Farag, A. A. and S. Y. Elhabian “Linear discriminant analysis (LDA).

A tutorial on data reduction”, www.cvip.uofl.edu/ wwwcvip/education/ECE523/LDA%20Tutorial.pdf [January 2011], 2008.

[19] M.Hariharan, L. S. Chee, O. C. Ai, and S. Yaacob , "Classification of Speech Dysfluencies Using LPC Based Parameterization Techniques", Journal of Medical Systems: 1-10, 2010.

[20] C.H. Park, H. Park, “A comparison of generalized linear discriminant analysis algorithms”, Pattern Recognition, vol. 41, pp. 1083-1097, 2008

[21] S. Bhattacharyya, A. Khasnobish, S. Chatterjee, A. Konar, and D. Tibarewala, "Performance analysis of LDA, QDA and KNN algorithms in left-right limb movement classification from EEG data," pp. 126-131.

[22] L. Zhiwei and S. Minfen, “Classification of Mental Task EEG Signals Using Wavelet Packet Entropy and SVM”, the Eighth International Conference on Electronic Measurement and Instruments, ‘ICEMI’, 2007, pp. 3-907 - 3-910.

[23] P. Refaeilzadeh, L. TANG, and H. LIU, "Cross-Validation," Arizona State University, 2008, pp. 1-6.

[24] A. Vincent, R. Shetty, M. Kurian, and B. K. Shetty, "Prospective, cross-sectional study, demonstrating efficacy of blue fixation target while recording Pattern Visual Evoked Potential in optic neuropathy," Documenta ophthalmologica, vol. 119, pp. 89-99, 2009.


186

[ieee 2011 ieee international conference on control system, computing and engineering (iccsce) -...

Documents