ORIGINAL ARTICLE

Mortality prediction of rats in acute hemorrhagicshock using machine learning techniques

Kyung-Ah Kim • Joon Yul Choi • Tae Keun Yoo •

Sung Kean Kim • KilSoo Chung • Deok Won Kim

Received: 29 January 2013 / Accepted: 8 June 2013 / Published online: 23 June 2013

� International Federation for Medical and Biological Engineering 2013

Abstract This study sought to determine a mortality

prediction model that could be used for triage in the setting

of acute hemorrhage from trauma. To achieve this aim,

various machine learning techniques were applied using

the rat model in acute hemorrhage. Thirty-six anesthetized

rats were randomized into three groups according to the

volume of controlled blood loss. Measurements included

heart rate (HR), systolic and diastolic blood pressures (SBP

and DBP), mean arterial pressure, pulse pressure, respira-

tory rate, temperature, blood lactate concentration (LC),

peripheral perfusion (PP), shock index (SI, SI = HR/SBP),

and a new hemorrhage-induced severity index (NI,

NI = LC/PP). NI was suggested as one of the good can-

didates for mortality prediction variable in our previous

study. We constructed mortality prediction models with

logistic regression (LR), artificial neural networks (ANN),

random forest (RF), and support vector machines (SVM)

with variable selection. The SVM model showed better

sensitivity (1.000) and area under curve (0.972) than the

LR, ANN, and RF models for mortality prediction. The

important variables selected by the SVM were NI and LC.

The SVM model may be very helpful to first responders

who need to make accurate triage decisions and rapidly

treat hemorrhagic patients in cases of trauma.

Keywords Hemorrhagic shock � Rat � Mortality �Machine learning � Support vector machine

1 Introduction

The World Health Organization (WHO) reported that

approximately 5 million people died from injury around the

world in 2008 [30]. Hemorrhagic shock accounts for 30 %

of these deaths [15]. Hemorrhagic shock is a clinical syn-

drome resulting from circulatory dysfunction that leads to

decreased tissue oxygenation, the accumulation of oxygen

debt, and ultimately to multiple organ system failure if left

untreated [29]. Thirty-three to 56 % of civilian deaths and

80–90 % of combat deaths occur during the pre-hospital

period [17]. The ability of a first responder to prioritize

patients and perform proper treatment has been emphasized

in instances of both civilian and military trauma [3, 16].

However, because there are few obvious symptoms in

patients, it is difficult to make triage decisions rapidly and

accurately and to utilize appropriate interventions early in

the decision-making process [14].

In the clinical setting, conventional vital signs such as

heart rate (HR), blood pressure (BP), pulse pressure (PPR),

and respiratory rate (RR) have been used to predict hem-

orrhagic shock severity based on the American College of

Surgeons Advanced Trauma Life Support (ATLS) for

Doctors (8th Student Course Manual) [2]. In addition, the

K.-A. Kim

Department of Biomedical Engineering, Chungbuk National

University, College of Medicine, Cheongju, Korea

J. Y. Choi � S. K. Kim � D. W. Kim (&)

Department of Medical Engineering, Yonsei University

College of Medicine, Seoul, Korea

e-mail: kdw@yuhs.ac

T. K. Yoo

Department of Medicine, Yonsei University, Seoul, Korea

S. K. Kim � D. W. Kim

Graduate Program in Biomedical Engineering,

Yonsei University, Seoul, Korea

K. Chung

Department of Electrical System, DongYang Mirae University,

Seoul, Korea

123

Med Biol Eng Comput (2013) 51:1059–1067

DOI 10.1007/s11517-013-1091-0

shock index (SI), defined as a ratio of HR to systolic blood

pressure (SBP), was a better measure of acute blood loss

than HR, SBP, or diastolic blood pressure (DBP) alone [4,

21]. Recent studies have concentrated on tissue status as

measured by lactate concentration (LC) and peripheral

perfusion (PP) to diagnose hemorrhagic shock severity [10,

29]. Our previous study using logistic regression (LR)

analyses demonstrated that a newly proposed hemorrhage-

induced severity index (NI), defined as the LC/PP ratio,

better predicted mortality in rats experiencing acute hem-

orrhage than SI as well as LC or PP alone [6]. However, an

artificial intelligence model is a very suitable alternative to

accurately predict hemorrhagic shock severity using vari-

ous physiological variables simultaneously.

The machine learning techniques have been applied to

diagnose some diseases such as acute appendicitis [12].

Recently, many studies have applied machine learning

techniques to the severity prediction model for hemor-

rhagic shock. Roberts et al. [23] used a polynomial neural

network in a swine model to predict mortality in hemor-

rhagic shock using SBP and mean arterial pressure (MAP).

Eftekhar et al. [11] compared logistic regression (LR) and

artificial neural network (ANN) models to predict mortality

from head trauma in humans and showed that ANN sig-

nificantly outperformed LR in prediction accuracy. Jang

et al. [14] recently applied an artificial neural network

(ANN) and support vector machine (SVM), and Choi et al.

[5] applied random forest (RF) classifiers to mortality

prediction models using a hemorrhagic rat model [5, 14].

Although each machine learning technique could predict

the severity of hemorrhagic shock, few studies simulta-

neously considered all of the aforementioned physiological

variables in constructing a prediction model. Furthermore,

few studies performed variable selection to achieve an

optimized model or compared performance among the

machine learning techniques in predicting mortality from

hemorrhagic shock. In this study, we validated the per-

formance of the machine learning techniques, LR, ANN,

RF, and SVM, and aimed to select an optimal model that

accurately predicts hemorrhagic shock mortality in the rat

model. Essential variables were selected from all physio-

logical variables before constructing each optimal model.

The models were then compared for sensitivity, specificity,

accuracy, and area under the curve (AUC) of the receiver

operating characteristic (ROC).

2 Methods

2.1 Experimental protocol [6]

Thirty-six male Sprague–Dawley rats (Orient, Seongnamsi,

Korea), weighing 300–400 g each, were divided into three

groups with 12 rats in each group depending on controlled

blood volume loss. After anesthesia with an isoflurane

inhalation system (RC2, VetEquip, Pleasanton, CA), blood

volumes of 2.0, 2.5, or 3.0 mL 100 g-1 were drawn

through a right carotid artery catheter using a withdrawal

syringe pump (Pump 11 Plus, Harvard Apparatus, South

Natick, MA) over 15 min for all groups (Fig. 1). The

average percentages of calculated total blood volume

withdrawn from these groups were 32.8, 41.0, and 49.2 %,

respectively [total blood volume (mL) = body weight

(g) 9 0.061] [31].

Uncontrolled hemorrhage was performed by amputation

of the tail at 75 % of its length at 1 min after initiating

volume-controlled hemorrhage. Tail blood was immedi-

ately directed into a tube containing heparinized saline, and

the blood volume collected was determined by weighing

the tube. The stump of the tail was tightly bandaged to stop

uncontrolled bleeding at the end of the controlled hemor-

rhage set at 30 min (Fig. 1). Since the major causative

factor in almost half of hemorrhage-related deaths occurs

during the acute period less than 2 h after injury [19], all

data were collected until death or through 150 min,

whichever came first [6, 19, 22]. Rats that survived

150 min were killed, and death was defined as the absence

of measured MAP [27]. All experimental procedures and

protocols used in this study were approved by the Institu-

tional Animal Care and Use Committee of Yonsei Uni-

versity Health System, Seoul, Korea.

2.2 Data acquisition [6]

The physiological variables such as HR, SBP, DBP, MAP,

PPR, RR, and TEMP were measured with a sampling

frequency of 1 kHz using LabChart 6 Pro software (AD

Instruments, Colorado Springs, CO). The data were fed

into an Analog/Digital conversion system (PowerLab 8/30,

AD Instruments). HR was obtained via ECG signals using

three invasive needle electrodes, placed in the left, right

forelegs, and the right hind leg, with a biological amplifier

(ML136, AD Instruments). An invasive blood pressure

amplifier was constructed using an instrumentation ampli-

fier (AD620, Analog Device, Norwood, MA) and a piezo-

resistive silicon pressure sensor (1620, Measurement

Specialist, Hampton, VA) that was connected to a right

femoral artery catheter. Respirations were assessed by

measuring changes in the thoracic cross-sectional area

using a respiratory belt (MLT1132, AD Instruments). We

measured temperature using a rectal probe (MLT1403, AD

Instruments), which was inserted approximately 5 cm into

the rectum. The SI was calculated dividing the measured

HR by SBP values.

Peripheral perfusion, defined as peripheral tissue perfu-

sion and skin microcirculation, was monitored continuously

1060 Med Biol Eng Comput (2013) 51:1059–1067

123

using a laser Doppler perfusion monitor (PeriFlux system

5000, Perimed, Sweden) with a probe (Probe 407, Peri-

med), which was attached to the right front sole of the paw.

The data were acquired with a sampling frequency of

32 Hz and analyzed by computer software (PeriSoft for

Windows, Perimed). Blood was sampled repetitively from

the left femoral vein using a 31-gauge insulin syringe (BD

Ultra-Fine, Becton–Dickinson, Franklin Lakes, NJ) for LC

measurement and then analyzed immediately with a

portable blood lactate analyzer (Lactate Pro LT-1710,

ARKRAY Inc., Kyoto, Japan), as shown in Fig. 1. Interim

LC values were calculated for every 10-s period between

blood draws through linear interpolation using computer

software (LabVIEW 2009, National Instruments, Austin,

TX) [23]. NI was calculated dividing the measured LC by

PP values.

As all measured data were averaged across every 10-s

period, six data sets were obtained per minute. We

analyzed data based on the changes between at the resting

state (100 %) for 30 s and immediately after the cessation

of bleeding for 10 s, as shown in Fig. 1. The reason for

analyzing over a short time period of 10 s was to limit the

effect of hemostasis as much as possible. These data points

were selected in order to mimic the scenario in which first

responders initially encounter injured and bleeding patients

who have lost unknown quantities of blood due to acute

hemorrhage. The physiological variables used for the

models were compared between the survival and death

groups using the Mann–Whitney U test with a significance

of 0.05 using SPSS version 18.0 (SPSS Inc., Chicago, IL)

as shown in Table 1.

2.3 Machine learning techniques

Logistic regression is used to generate a predictive model

for dichotomous response variables by fitting available data

to a logistic function, which always takes on values

between zero and one. The logistic function consists of

regression coefficients with the respective input variables.

Each of the regression coefficients describes the size of the

contribution of that risk factor. LR models are typically the

gold standard among medical researchers for comparison

in machine learning studies [12]. We compared the per-

formance of the LR model to that of the ANN, RF, and

SVM models.

The concept of ANN is to imitate a biological neural

network. In this study, the architecture of the ANN con-

sisted of three layers: one input, one hidden, and one out-

put. Each layer of the network has elementary processing

units called nodes. There were 11 input nodes and one

output node for survival or death. In general, the perfor-

mance of ANN models relies on the number of hidden

nodes. Since the optimal number of hidden nodes is

determined by empirical method [27], this study trained the

networks that were applied from two to ten hidden nodes to

obtain an optimal ANN model. The ANN algorithm used in

this study was based on the Levenberg–Marquardt back-

propagation algorithm with epoch of 100 and gradient

of\e-50. Back-propagation is an algorithm that calculates

the minimum error by changing weight of layers.

Resting Bleeding Post bleeding

0 1516 20 25

3035 40 45 75 135105

150 (min)

Tail amputation: Lactate concentration measurement

Analysis

10’’30’’14

Fig. 1 Experimental protocol including measurement of lactate concentration, tail amputation, and period of data analysis [6]

Table 1 Comparison of all measured data [mean (standard devia-

tion), p value] at the completion of hemorrhage versus at rest (100 %)

for the survival and death groups [6]

Parameter [%] Survival (n = 16) Death (n = 20) p value

Total bleeding 42.3 (5.1) 51.2 (4.4) \0.001*

NI; LC/PP 437 (196) 1,654 (1,297) \0.001*

SI; HR/SBP 213 (72) 457 (143) \0.001*

PP [PU] 34 (12) 14 (9) \0.001*

SBP [mmHg] 47 (20) 23 (9) \0.001*

MAP [mmHg] 43 (16) 22 (9) \0.001*

PPR [mmHg] 65 (40) 24 (13) \0.001*

DBP [mmHg] 39 (16) 22 (9) \0.001*

TEMP [�C] 98 (1) 97 (1) 0.001*

LC [mmol L-1] 132 (40) 181 (54) 0.004*

RR [bpm] 81 (16) 64 (22) 0.013*

HR [bpm] 90 (12) 95 (9) 0.080

Total bleeding [%] = (bleeding/total blood volume) 9 100

NI new index, SI shock index, PP peripheral perfusion, PU perfusion

unit, SBP systolic blood pressure, MAP mean arterial pressure, PPR

pulse pressure, DBP diastolic blood pressure, TEMP temperature, LC

lactate concentration, RR respiration rate, HR heart rate

* p \ 0.05

Med Biol Eng Comput (2013) 51:1059–1067 1061

123

Random forest is an ensemble classification that consists

of many independent decision trees with outputs. Each tree

is grown independently in combination with a bagging idea

and a random selection of input variables for the condi-

tions. The final result is based on a majority vote of all trees

[12]. The goal of RF is to accurately classify by controlling

the number of the trees. We investigated mortality pre-

diction models using trees of 50, 100, 200, 300, 400, and

500 to find the optimal RF model.

Support vector machines maps data to a higher dimen-

sional space through a kernel function to linearly separate

data patterns. The data are separated into two groups by the

training data referred to as a support vector. SVM models are

determined by choosing the maximum-margin hyperplane

with the nearest support vector of two groups [7]. SVM

improves the accuracy of a model through the optimization

of separating space using the kernel function. To obtain the

optimal SVM model, we carried out exhaustive grid search

by changing penalty parameter C of 0.1, 1, and 10 and scaling

factor r of 0.1 and 1 using the Gaussian kernel function.

2.4 Model construction

The data set in this study was obtained from our previous

study [6]. For model construction, 11 variables including

HR, SBP, DBP, MAP, PPR, TEMP, RR, LC PP, SI, and NI

were used for input data sets. This study obtained one data

set from each rat. Therefore, 36 (1 set/10 s 9 36 rats) data

sets were obtained. The numbers for the survival and death

groups were 16 and 20, respectively. The 36 data sets were

divided randomly into two mutually exclusive data sets. Of

these, 66.7 % (n = 24; survival = 10, death = 14) were

used as a training set to select the optimal variables and

construct the models, and the remaining 33.3 % (n = 12;

survival = 6, death = 6) were used as a testing set. Before

constructing the models, we adopted a grid search in which

a range of parameter values, mentioned above in Sect. 2.3,

for each model was tested to obtain the optimal classifi-

cation using tenfold cross-validation. The optimal ANN

was set with three neurons within the hidden layer. The

optimal RF used 100 trees, and the number of predictors for

each node was set to default. For the SVM, a Gaussian

kernel function with a penalty parameter C of 0.1 and a

scaling factor r of 1 was set.

In the upper part of Fig. 2, we demonstrate the variable

selection with the grid search to minimize and optimize the

LR, ANN, RF, and SVM models. This variable selection

could be helpful for first responders to reduce the impres-

sion time needed on scene in cases of injured bleeding

patients. All of the physiological variables, including HR,

SBP, DBP, MAP, PPR, TEMP, RR, LC, PP, SI, and NI,

were considered for variable selection. For variable selec-

tion, backward elimination using the priority of the

variables was used. Priority of the variables was deter-

mined based on the embedded method for each machine

learning technique [12, 24]. Subsequently, tenfold cross-

validation was repeated to calculate the mean accuracy of

each cross-validation process by progressively eliminating

the least contributing ones until only the most essential

variables remained (backward elimination) [8]. The highest

ranked variables with the best cross-validated accuracy

were chosen as the optimal variables.

The prediction model was constructed using the optimal

variables, and then, its performance was evaluated using

the remaining testing set. To compare the performance of

the mortality prediction models with the selected variables,

we obtained sensitivity, specificity, and accuracy. We also

drew an ROC curve for each model and calculated the area

under the ROC curve (AUC). ANN, RF, and SVM were

analyzed using MATLAB Version 2010 a (Mathworks

Inc., Natick, MA), and LR was analyzed using SPSS 18.0.

3 Results

Table 1 shows some results of our previous study that total

bleeding volume (%) and the percent changes in all

Data set : physiological variables from rats

Training set / testing set randomly (24 / 12)

Priority determinationamong variables with embedded method

Repeat 10-fold cross validation with progressive elimination of variables contributing the least (backward elimination)

Input variables 10-fold cross validation Mean accuracy

1aV11V1 V2 V10V3

V11V2 V10V3

V11V10

V11

2a

10a

11a

Optimal variables determination by highest mean accuracy

Mortality prediction model constructionwith the selected variables

Optimal model evaluation(sensitivity, specificity, accuracy, AUC)

In training set

In testing set

Fig. 2 Flowchart of backward elimination variable selection, mor-

tality prediction model construction, and performance validation

1062 Med Biol Eng Comput (2013) 51:1059–1067

123

measured variables (mean ± standard deviation) at the

completion of hemorrhage versus at rest (100 %) for both

the survival and death groups [6]. NI, SI, and LC for the

survival and death groups changed much more than any

other variable, suggesting that they were more sensitive to

acute hemorrhage for 15 min than other variables. The

decreases in PP and BPs including SBP, MAP, PPR, and

DBP were relatively greater than those of TEMP, RR, and

HR. There were significant differences between the two

groups for total bleeding volume and all variables

(p \ 0.05), except for HR.

Figure 3 shows the relative changes (%) in HR, SBP, SI,

LC, PP, and NI from 1 min before hemorrhage to 5 min

after hemorrhage cessation versus at rest [6]. The above six

variables were selected because SI and NI including their

dependent variables, HR, SBP, LC, and PP were consid-

ered important. As LC increased and PP decreased quickly

during hemorrhage, the relative LC/PP ratio increased

considerably, by more than 400 % in the survival group

and greater than 1,600 % in the death group at the end of

hemorrhage (Fig. 3). These considerable changes as a

result of hemorrhage suggest NI to be an attractive prog-

nostic tool for surviving hemorrhagic shock. The signifi-

cant differences between the survival and death groups

were obtained by Mann–Whitney U test [6].

Using variable selection from 11 variables in each

machine learning technique, only the selected variables in

each technique are shown in Table 2. SI, TEMP, and LC

were selected for the LR model, while SI, LC, and HR

were selected for the ANN model. Only SI was selected for

the RF model, while NI and LC were selected for the SVM.

It was unexpected results that TEMP and HR were selected

variables for the LR and ANN model, because their P

values were relatively larger than those of other variables.

Table 3 lists the sensitivity, specificity, accuracy, and AUC

for the LR, ANN, RF, and SVM models. Of the four

0

30

60

90

120

150

14 18 22 26 30 34

SBP

(%) * * ************************

******

Time (min)

0

30

60

90

120

150

14 18 22 26 30 34

PP

(%) *

********************************

Time (min)

0

30

60

90

120

150

14 18 22 26 30 34

SurvivalDeath

HR

(%)

*** ******

Time (min)

Hemorrhage

0

100

200

300

400

14 18 22 26 30 34

LC

(%)

* ******

****

****

**

Time (min)

0

100

200

300

400

500

600

14 18 22 26 30 34

SI (%

)

* **********

****

****

**************

Time (min)

0

1000

2000

3000

4000

5000

6000

14 18 22 26 30 34

NI

(%)

* ********

****

****

****

****

**

**

Time (min)

Fig. 3 Relative changes in HR,

SBP, SI, LC, PP, and NI during

hemorrhage [6]. All values are

percentages of the resting

values, which were designated

as 100 %. Error bars indicate

standard deviation. *p \ 0.05,

**p \ 0.01. HR heart rate, SBP

systolic blood pressure, SI shock

index, LC lactate concentration,

PP peripheral perfusion, NI new

index

Med Biol Eng Comput (2013) 51:1059–1067 1063

123

models, the SVM model showed the best sensitivity (1.000)

and AUC (0.972). Therefore, NI and LC, which were

selected for the SVM, are expected to be important factors

in predicting mortality from acute hemorrhage. Fig. 4

depicts the ROC curves for LR, ANN, RF, and SVM and

agrees with the AUCs in Table 3.

Figure 5 shows the distribution of the survival and death

groups of the testing set on the two-dimensional space of

LC and NI. The curve of the SVM model constructed by

the training set is depicted on the space with a nonlinear

boundary line between the survival and death groups. The

boundary line was determined by ROC analysis. The two-

dimensional feature space may be used simply to predict

severity when medical staff or first responders encounter

patients with hemorrhage.

4 Discussion

Variable selection methods have been conducted in many

studies to optimize predictive models and to determine risk

factors for certain diseases [12, 24]. This study applied var-

iable selection to the prediction of acute hemorrhagic shock

severity, not only to obtain optimal mortality prediction

models, but also to determine the influential physiological

variables. There are few studies investigating conventional

vital signs and recent diagnostic indices like LC and PP at the

same time using various machine learning techniques. In this

study, variables including NI and LC were selected in the

optimal SVM model as shown in Table 2. NI had higher

priority than LC in priority determination applying an

embedded method. Although this study is different from our

previous study, which used LR, in which no variable selec-

tion was performed, the previous study also showed that NI

had higher AUCs for predicting morality and higher corre-

lation with mortality than others [6]. Therefore, NI could be

helpful for first responders to rapidly and accurately predict

hemorrhagic shock severity. Human studies are warranted to

confirm that NI has a higher correlation with hemorrhagic

shock severity than other variables.

Variables are often strongly inter-correlated when per-

forming variable selection for solving clinical problems. It

is well known that conventional logistic regression tends to

Table 2 Variable selection in machine learning for mortality pre-

diction models

Variable LR ANN RF SVM

NI x

SI x x x

TEMP x

LC x x x

HR x

LR logistic regression, ANN artificial neural network, RF random

forest, SVM support vector machine, NI new index, SI shock index,

TEMP temperature, LC lactate concentration, HR heart rate

Table 3 Performance of the LR, ANN, RF, and SVM mortality

prediction models

LR ANN RF SVM

Sensitivity 0.678 0.833 0.833 1.000

Specificity 1.000 1.000 1.000 0.833

Accuracy 0.833 0.917 0.917 0.917

AUC 0.833 0.917 0.903 0.972

LR logistic regression, ANN artificial neural network, RF random

forest, SVM support vector machine, AUC area under curve

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

LR

ANN

RF

SVM

Sens

itivi

ty

1-Specificity

Fig. 4 ROC curves of LR, ANN, RF, and SVM models. LR logistic

regression, ANN artificial neural network, RF random forest, SVM

support vector machine

0

1000

2000

3000

4000

5000

6000

7000

0 50 100 150 200 250 300 350

Death group Survival group SVM boundary line

New

inde

x (%

)

Lactate concentration (%)

Fig. 5 Two-dimensional feature space for LC and NI with distribu-

tion of the survival and death groups. The curve represents a boundary

of the SVM model

1064 Med Biol Eng Comput (2013) 51:1059–1067

123

be over-fitting and there is coefficient instability since it

obtains variables by choosing the more important variables

and disposing of the less important variables when the

variables are strongly inter-correlated [28]. Recently, many

studies in bioinformatics have suggested the adequate use

of all variables deemed important in order to improve

accuracy without eliminating variables [33]. In this study,

the use of NI and LC for the SVM model showed better

performance even though NI and LC were inter-correlated.

SVMs determine the important variables based on relative

order that can improve accuracy at the separating boundary

[20]. The double inclusion of LC in the SVM model during

variable selection was not twice as more important than PP

because the variables were selected depending on the rel-

ative order; however, this showed that LC was important

for developing the separating boundary when considering

all variables. Since the importance of the SVM model is

only involved in developing the boundary, it is not guar-

anteed that LC is highly correlated with survival and death.

The ANN, RF, and SVM models showed greater accu-

racy and AUC than the LR model, as shown in Table 3.

Although LR has been used as the gold standard to com-

pare the diagnostic prediction models suggested in many

machine learning studies, the ANN, RF, and SVM models

in this study were preferable as mortality prediction models

resulting from acute hemorrhagic shock in terms of per-

formance. It is to be noted that only the SVM model

showed 100 % sensitivity. For our purpose, sensitivity was

defined as the correctly identified probability of death in

rats. Although the ANN, RF, and SVM models showed the

same accuracy, the SVM model showed the greatest AUC.

Generally, AUC is known as a better predictor than accu-

racy in evaluating learning algorithms [13]. It is therefore

reasonable that the SVM model was superior to the LR,

ANN, and RF models in terms of sensitivity and AUC.

Logistic regression is widely used in biomedicine as a

prediction model. However, the LR model showed the

worst performance in predicting mortality in this study. LR

models use linear combinations of variables and, therefore,

are not adept at modeling nonlinear complex interactions,

whereas ANN models are rich and flexible nonlinear sys-

tems [11]. ANN has a disadvantage of a local minima

problem when creating models. However, the SVM model

can create a model that converges to a global minimum

because it is based on structural risk minimization [14].

The RF model showed relatively poor performance. Since

RFs construct models based on decision trees, the models

have rectangular partitioning of the input space when

capturing a linear boundary between groups. This staircase

approximation leads to an inaccurate fit of the linear or

complex boundaries [26].

Support vector machines, especially with the Gaussian

kernel function, is well known for its high performance as a

machine learning technique in predicting diseases [32]. In

this study, the SVM model with the Gaussian kernel

function performed best in predicting mortality due to

acute hemorrhagic shock in the rat. The SVM model

considers input variables based on their own characteristics

of nonlinearity and high dimension. The SVM model dealt

with a separating space consisting of variables in high

dimension and was thus able to consider all variables,

improving its performance in predicting hemorrhagic

shock severity [1]. One of the disadvantages of SVM was

that it required many trials to construct an optimal SVM

model in comparison with other machine learning tech-

niques. Since there is no reliable method for selecting the

optimal penalty parameter, C, and scaling factor, r, of the

Gaussian kernel function, we carried out an exhaustive grid

search by changing C and r.

This study showed the potential of the SVM model that

used NI and LC in helping first responders predict the

severity of hemorrhage in the field. It took only 0.15 s for

the SVM model to decide whether a rat belonged to the

death or survival group using the selected variables;

therefore, the processing time is not a problem. In the

current field, clinical first responders or paramedics rely on

their subjective judgment to predict the severity of hem-

orrhage in patients [16]. This study suggested an objective

index to predict the severity with high accuracy, using the

SVM model based on NI and LC. Even though PP and LC

obtained by a portable device cannot be measured in the

field at present, our previous study showed good respon-

siveness of PP to acute bleeding as shown in Fig. 3 [6].

The portable lactate analyzers used in this study enabled

measurements within 60 s, and exhibited high correlation

with traditional serum LC measurement without profes-

sional skills [6, 29]. Therefore, the SVM model with a short

processing time could provide more objective judgment to

clinical first responders without special skills by not only

prioritizing patients in battle fields or mass casualty situa-

tions, but also when performing more intense interventions

for patients in instances of civilian trauma. It would also

satisfy the need for rational transport distance for patients

[25] and suitable preparation for the patients rapidly in the

trauma center. In other words, more effective medical care

could be offered for patients with hemorrhage in extreme

time pressured situations. However, further work with NI

and the effects of resuscitation on the SVM model is

warranted.

There were some limitations to this study. First, our

experiment was performed under anesthesia. As anesthesia

might affect the cardiovascular system [18], we used the

isoflurane inhalation system instead of venous anesthetic

injection to minimize the effects on the cardiovascular

system. Under isoflurane inhalation anesthesia, the physi-

ological responses of an animal model are more reliable

Med Biol Eng Comput (2013) 51:1059–1067 1065

123

because of its minimal metabolism, biotransformation, and

excretion [9]. Second, a validation study for humans was

not performed. In emergency situations, one would not be

provided with resting variable measurements. Therefore,

we suggest to use resting variables with normally distrib-

uted data ranges for humans [6].

In conclusion, the SVM model performed best in terms

of sensitivity and AUC in comparison with LR and other

machine learning techniques such as ANN and RF for the

rat model in acute hemorrhagic shock. Even though the

SVM model is not applicable to humans at present, this

study showed its potential as a promising technique for

clinical application. If confirmation of this SVM model

were to be achieved in humans, this technique could help

first responders predict lethal hemorrhagic shock in cases

of trauma.

Acknowledgments This study was supported by a faculty research

Grant from Yonsei University College of Medicine for 2011 (6-2011-

0087).

References

1. Agranoff D, Fernandez-Reyes D, Papadopoulos MC, Rojas SA,

Herbster M, Loosemore A, Tarelli E, Sheldon J, Schwenk A,

Pollok R, Rayner CF, Krishna S (2006) Identification of diag-

nostic markers for tuberculosis by proteomic fingerprinting of

serum. Lancet 368(9540):1012–1021

2. American College of Surgeons (2008) ATLS: advanced trauma

life support for doctors (student course manual), 8th edn. Am Coll

of Surg, Chicago

3. Benedek DM, Fullerton C, Ursano RJ (2007) First responders:

mental health consequences of natural and human-made disasters

for public health and public safety workers. Annu Rev Public

Health 28:55–68

4. Birkhahn RH, Gaeta TJ, Terry D, Bove JJ, Tloczkowski J (2005)

Shock index in diagnosing early acute hypovolemia. Am J Emerg

Med 23(3):323–326

5. Choi JY, Kim SK, Lee WH, Yoo TK, Kim DW (2012) A survival

prediction model of rats in hemorrhagic shock using the random

forest classifier. Proceedings of the 34th International Conference

of the IEEE Engineering in Medicine and Biology Society, San

Diego, IEEE Service Center, Piscataway, NJ, pp 5570–5573. doi:

10.1109/IEMBS.2012.6347256

6. Choi JY, Lee WH, Yoo TK, Park I, Kim DW (2012) A new

severity predicting index for hemorrhagic shock using lactate

concentration and peripheral perfusion in a rat model. Shock

38(6):635–641

7. Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn

20(3):273–297

8. Dash M, Liu H (2003) Consistency-based search in feature

selection. Artif Intell 151(1–2):155–176

9. Davis JA (2008) Mouse and rat anesthesia and analgesia. Curr

Protoc Neurosci (Appendix 4: Appendix 4B). doi:10.1002/0471

142301.sna04bs42

10. Dutton RP (2007) Current concepts in hemorrhagic shock.

Anesthesiol Clin 25(1):23–34

11. Eftekhar B, Mohammad K, Ardebili HE, Ghodsi M, Ketabchi E

(2005) Comparison of artificial neural network and logistic

regression models for prediction of mortality in head trauma

based on initial clinical data. BMC Med Inform Decis Mak 5:3

12. Hsieh CH, Lu RH, Lee NH, Chiu WT, Hsu MH, Li YC (2011)

Novel solutions for an old disease: diagnosis of acute appendicitis

with random forest, support vector machines, and artificial neural

networks. Surgery 149(1):87–93

13. Huang J, Ling CX (2005) Using AUC and accuracy in evaluating

learning algorithms. IEEE Trans Knowl Data Eng 17(3):299–310

14. Jang KH, Yoo TK, Choi JY, Nam KC, Choi JL, Kwon MK, Kim

DW (2011) Comparison of survival predictions for rats with

hemorrhagic shocks using an artificial neural network and support

vector machine. Proceedings of the 33rd International Conference

of the IEEE Engineering in Medicine and Biology Society,

Boston, IEEE Service Center, Piscataway, NJ, pp 91–94 doi:

10.1109/IEMBS.2011.6089904

15. Kauvar DS, Lefering R, Wade CE (2006) Impact of hemorrhage

on trauma outcome: an overview of epidemiology, clinical pre-

sentations, and therapeutic considerations. J Trauma 60(Suppl.

6):S3–S11

16. Lerner EB, Schwartz RB, Coule PL, Weinstein ES, Cone DC,

Hunt RC, Sasser SM, Liu JM, Nudell NG, Wedmore LS, Ham-

mond J, Bulger EM, Salomone JP, Sanddal TL, Markenson D,

O0Connor RE (2008) Mass casualty triage: an evaluation of the

data and development of a proposed national guideline. Disaster

Med Public Health Prep 2(Suppl 1):S25–S34

17. Letson HL, Dobson GP (2011) Unexpected 100% survival fol-

lowing 60% blood loss using small-volume 7.5% NaCl with

adenocaine and Mg (2?) in the rat model of extreme hemorrhagic

shock. Shock 36(6):586–594

18. Lomas-Niera JL, Perl M, Chung CS, Ayala A (2005) Shock and

hemorrhage: an overview of animal models. Shock 24(Suppl 1):

33–39

19. Macias CA, Kameneva MV, Tenhunen JJ, Puyana JC, Fink MP

(2004) Survival in a rat model of lethal hemorrhagic shock is

prolonged following resuscitation with a small volume of a

solution containing a drag-reducing polymer derived from aloe

vera. Shock 22(2):151–156

20. Mukkamala S, Janoski G, Sung A (2002) Intrusion detection

using neural networks and support vector machines. Proceedings

of the 2002 International Joint Conference on Neural Networks,

pp 1702–1707. doi:10.1109/IJCNN.2002.1007774

21. Nakasone Y, Ikeda O, Yamashita Y, Kudoh K, Shigematsu Y,

Harada K (2007) Shock index correlates with extravasation on

angiographs of gastrointestinal hemorrhage: a logistics regression

analysis. Cardiovasc Intervent Radiol 30(5):861–865

22. Not LG, Marchase RB, Fulop N, Brocks CA, Chatham JC (2007)

Glucosamine administration improves survival rate after severe

hemorrhagic shock combined with trauma in rats. Shock

28(3):345–352

23. Roberts DA, Holcomb JB, Parker BE Jr, Sondeen JL, Pusateri

AE, Brady WJ Jr, Sweenor DE, Young JS (2002) The use of

polynomial neural networks for mortality prediction in uncon-

trolled venous and arterial hemorrhage. J Trauma 52(1):130–135

24. Saeys Y, Inza I, Larranaga P (2007) A review of feature selection

techniques in bioinformatics. Bioinforma 23(19):2507–2517

25. Schull MJ, Morrison LJ, Vermeulen M, Redelmeier DA (2003)

Emergency department overcrowding and ambulance transport

delays for patients with chest pain. CMAJ 168(3):277–283

26. Statnikov A, Wang L, Aliferis CF (2008) A comprehensive

comparison of random forests and support vector machines for

microarray-based cancer classification. BMC Bioinformatics

9:319. doi:10.1186/1471-2105-9-319

27. Takasu A, Sakamoto T, Okada Y (2006) Effect of induction rate

for mild hypothermia on survival time during uncontrolled

hemorrhagic shock in rats. J Trauma 61(6):1330–1335

1066 Med Biol Eng Comput (2013) 51:1059–1067

123

28. Waldron L, Pintilie M, Tsao MS, Shepherd FA, Huttenhower C,

Jurisica I (2011) Optimized application of penalized regression

methods to diverse genomic data. Bioinformatics 27(24):3399–

3406

29. Wilson M, Davis DP, Coimbra R (2003) Diagnosis and moni-

toring of hemorrhagic shock during the initial resuscitation of

multiple trauma patients: a review. J Emerg Med 24(4):413–422

30. World Health Organization (2010) World health statistics 2010.

Switzerland, World Health Organization Press, Geneva, pp 62–70

31. Yang S, Zou LY, Bounelis P, Chaudry I, Chatham JC, Marchase

RB (2006) Glucosamine administration during resuscitation

improves organ function after trauma hemorrhage. Shock

25(6):600–607

32. Zhao ZD, Lou YY, Ni JH, Zhang J (2012) RBF-SVM and its

application on reliability evaluation of electric power system

communication network. In: Machine Learning Cybernetics

(ICMLC), 2012 Annual International Conference of the IEEE,

pp.1188–1193. doi: 10.1109/ICMLC.2009.5212365

33. Zou H, Hastie T (2008) Regularization and variable selection

via the elastic net. J R Stat Soc Series B Stat Methodol

67(2):301–320

Med Biol Eng Comput (2013) 51:1059–1067 1067

123