Posted: March 12th, 2023

Summary of a literature review section

Can I have the literature review section summarized in a page or 1.5 page?

Accepted Manuscript

Classification of Transcranial Doppler Signals using Individual and
Ensemble Recurrent Neural Networks

Manjeevan Seera , Chee Peng Lim , Kay Sin Tan ,
Wei Shiung Liew

PII: S0925-2312(17)30568-4
DOI: 10.1016/j.neucom.2016.05.117
Reference: NEUCOM 18281

To appear in: Neurocomputing

Received date: 2 November 2015
Revised date: 22 April 2016
Accepted date: 15 May 2016

Please cite this article as: Manjeevan Seera , Chee Peng Lim , Kay Sin Tan , Wei Shiung Liew ,
Classification of Transcranial Doppler Signals using Individual and Ensemble Recurrent Neural Net-
works, Neurocomputing (2017), doi: 10.1016/j.neucom.2016.05.117

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.

http://dx.doi.org/10.1016/j.neucom.2016.05.117

http://dx.doi.org/10.1016/j.neucom.2016.05.117

ACCEPTED MANUSCRIP

T

ACCEPTED M
ANUSCRIP

T

Classification of Transcranial Doppler Signals using Individual and Ensemble Recurrent Neural

Networks

Manjeevan Seera a*, Chee Peng Lim b, Kay Sin Tan c, Wei Shiung Liew d
a Faculty of Engineering, Computing and Science, Swinburne University of Technology (Sarawak Campus), Malaysia
b Centre for Intelligent Systems Research, Deakin University, Geelong, Victoria, Australia
c University Malaya Medical Centre, University of Malaya, Kuala Lumpur, Malaysia
d Faculty of Computer Science and Information Technology, University of Malaya, Kuala Lumpur, Malaysia

Abstract Transcranial Doppler (TCD) is a reliable technique with the advantage of being non-invasive for the diagnosis of

cerebrovascular diseases using blood flow velocity measurements pertaining to the cerebral arterial segments. In this study,

the Recurrent Neural Network (RNN) is used to classify TCD signals captured from the brain. A total of 35 real,

anonymous patient records are collected, and a series of experiments for stenosis diagnosis is conducted. The extracted

features from the TCD signals are used for classification using a number of RNN models with recurrent feedbacks. In

addition to individual RNN results, an ensemble RNN model is formed in which the majority voting method is used to

combine the individual RNN predictions into an integrated prediction. The results, which include the accuracy, sensitivity,

and specificity rates as well as the area under the Receiver Operating Characteristic curve, are compared with those from the

Random Forest ensemble model. The outcome positively indicates the usefulness of the RNN ensemble as an effective

method for detecting and classifying blood flow velocity changes due to brain diseases.

Keywords: Doppler signals; recurrent neural network; pattern classification; stenosis.

1. Introduction

In the medical field, the Doppler principle is based on emitting sound waves into the body, and then monitoring the

changes in frequency between the signals transmitted and signals reflected back from the target, detected by a transducer

[1]. Satomura first used the Doppler technique in 1959 to measure Doppler noise (signals) from the blood vessels [2].

Because of the non-invasive nature, Doppler-based techniques became popular, and have been widely used since then.

During an ultrasound examination of the carotid arteries, Austen et al. [3] discovered transient increased in the Doppler

signal intensity. Similar signals were recorded in cases covering decompression sickness and angiography [4]. In [5],

Doppler techniques were deployed for measuring the blood flow velocity in diagnosing vascular diseases. Specifically,

shifts from the red blood cells detected by the Doppler technique were used for computing the blood velocity [5]. Indeed,

Doppler sonography is typically preferred since angiography is costly and is an invasive method [6].

The key principle of transcranial Doppler (TCD) is similar to that of Doppler ultrasound [7]. Introduced by Aaslid et al.

in 1982 [7], TCD sonography has been used to measure the blood flow velocity in large basal arteries in a non-invasive

manner. The TCD technique provides measurements of blood flow changes, which are valuable in determining the decrease

in cerebral blood flow that can cause neurologic sequela [2]. Because the TCD machine is portable, reliable, and cheaper as

compared with other test methods [2], it is widely used nowadays.

Studies of TCD are generally concerned with evaluating stenosis of intracranial arterial, cerebral vasospasm, cerebral

arteriovenous malformations, and cerebral hemodynamics [1]. Typical intra-brain blood flow measurements need high-end

imaging methods. However, the diagnosis can be made using TCD, by detecting the increased or decreased blood flow

velocity [1]. In the diagnosis of stenosis and occlusion of arteries, TCD is used for blood flow velocity measurements in the

brain arteries [1]. This can be an alternative diagnostic method as compared with angiography and magnetic resonance

methods [1]. Based on evidences from a variety of different cerebrovascular pathologies, TCD has been shown to be

practical for clinical settings [7].

Computer aided diagnosis has been seen as a preferred direction for disease diagnosis [8]. Among many different

methods, artificial neural networks (ANN) have been used in various applications, e.g. medical signal processing [9-10].

Application of ANNs to Doppler related problems includes classifying carotid arterial Doppler ultrasound signals [11-13].

In [11], the principal component analysis and fuzzy c-means clustering methods were used with a complex-valued artificial

neural network (CVANN) to classify carotid artery Doppler signals in the early phase of atherosclerosis. In [12], the fast

Fourier transform was used to process Doppler signals, and the processed signals were subsequently classified using the

CVANN model. In [13], discrete wavelet transform was used for signal pre-processing while wavelet transform-CVANN

and complex wavelet transform-CVANN models were used for classification of Doppler signals. Using TCD signals

measured from the brain temporal region, the k-nearest neighbour (k-NN) algorithm was used for classifying the recorded

signals [1]. The correlation dimensions with the maximum Lyapunov exponents were employed as the input features to the

k-NN algorithm.

ACCEPTED MANUSCRIPT

ACCEPTED M
ANUSCRIP

T

The main contribution of this paper is an analysis of TCD signals from a cohort of real patients. An ensemble of

Recurrent Neural Networks (RNNs) is used to classify various signals acquired from anonymous patients. RNNs have been

previously used in an intracranial pressure model [14], with the results analysed, discussed, and compared with those from

the Random Forest Ensemble (RFE) model [15]. The rest of this paper is organised as follows. A literature review is first

presented in Section 2. The RNN model is described in Section 3. In Section 4, the method used for TCD signal analysis is

detailed, and the experimental results and discussion are presented in Section 5. Concluding remarks and suggestions for

further work are given in Section 6.

2. Literature review

In this section, a review related to stenosis and Doppler signals is presented. A variety of methods are described, and a

summary is given in Table 1. Detection of ophthalmic artery (OA) Doppler signals for the Behcet disease was conducted in

[16]. Using Doppler signals from the OA, a spectral analysis was performed using the least-square autoregressive (AR)

method [16]. Based on the Multilayer Perceptron (MLP) network, classification rates of 96.43% and 93.75% were achieved

for healthy patients and patients with the Behcet disease, respectively [16]. The diagnosis of OA and internal carotid artery

(ICA) disorders was presented in [17]. Doppler signals of OA and ICA were first decomposed using the wavelet transform

and statistical features with time–frequency representations [17]. The MLP network trained with the Levenberg Marquart

(LM) algorithm was used for classification. The accuracy rates achieved were 96.43% for healthy patients, 97.44% for

patients with ICA stenosis, and 96.97% for patients with ICA occlusion.

The diagnosis of mitral heart valve stenosis using Doppler ultrasound signals was conducted in [18] and [19]. In [18],

Doppler signals were transformed into the Power Spectral Density (PSD). Using MLP trained with LM, the correct

classification rate achieved was 94% [18]. In [19], the short time Fourier transformation (STFT) method was used for

processing Doppler signals. Similarly, using the MLP trained with LM, an accuracy rate of 97.8% was obtained [19]. To

diagnose three heart valve disorders, i.e. aortic insufficiency, aortic stenosis, and pulmonary stenosis, a multi-resolution

wavelet-based algorithm was deployed to extract a set of statistical features pertaining to the phonocardiogram signals [20].

The Daubechies wavelet filter with five decomposition levels was then used. The classification accuracy achieved using

MLP with the backpropagation (BP) training algorithm was 94.42% [20].

The diagnosis of occlusive arterial disease through the analysis of femoral artery Doppler signals was presented in [21].

PSD of the signals was generated using Welch and AR modelling [21]. Using MLP trained with LM, the accuracy rates

reported were 98% to 99%, respectively [21]. Doppler signals from ICA were analysed in [22]. Specifically, spectral

analysis of the signals was performed using wavelet transform. For detection of stenosis and occlusion in ICA, MLP trained

with LM was employed [22]. Classification rates of 96%, 96.15%, and 96.30% were achieved for healthy patients, patients

with ICA stenosis, and patients with ICA occlusion, respectively [22]. Coronary arteriography of significant coronary

stenosis based on a national cardiac catheterization database was used in [23]. A total of eleven features were used as the

inputs to MLP [23]. The specificity and sensitivity rates of 26% and 100% were reported for patients without and with

significant stenosis, respectively [23].

Both MLP and radial basis function (RBF) networks were used in [24], [25], and [26] for undertaking medical diagnosis

tasks. Congestive heart failure and chronic obstructive pulmonary disease from patients were differentiated based on a total

of 42 clinical variables as a result of consultation with cardiologists [24]. MLP with Bayesian regularization produced

83.9% sensitivity and 86% specificity rates, while RBF with k-means clustering yielded 81.8% sensitivity and

88.4%

specificity rates [24]. On the other hand, PSD was used to process Doppler signals for classifying middle cerebral artery

stenosis, and classification rates of 94.2% and 88.4% were achieved by MLP and RBF, respectively [25].

Similar to [25], Doppler signals were used for classifying left and right ICA stenosis [26]. A number of features

including obstructed veins from the coroner angiography, intimal thickness, and plaque formation were used as the inputs to

both MLP and RBF networks [26]. Classification rates of 87.5% and 80% were achieved by MLP and RBF, respectively

[26]. The Lyapunov exponents were computed from Doppler ultrasound signals in [27]. The probabilistic

neural network

(PNN) was used to classify Doppler signals using the extracted features. The PNN accuracy rates were stable, ranging

between 98.17% and 98.18% [27]. The adaptive neuro-fuzzy inference system (ANFIS) classifier was used in [28] and

[29]. Automatic detection of OA stenosis was conducted in [28]. Two features, namely resistivity and pulsatility indices,

were fed into ANFIS for classification [28]. The best accuracy rate of 95.83% was achieved [28].

A classification system for stenosis using the Doppler signals from the mitral valve was presented in [29]. STFT was

used to extract features from Doppler signals [29]. The classification rate of ANFIS was 98% [29]. The right and left hand

Ulnar artery Doppler signals were utilized in [30]. Using the spectral analysis method of Multiple Signal Classification

(MUSIC), different features were extracted [30]. ANFIS was then used for classification, and the accuracy rates achieved

ranged from 91.25% to 95% [30].

Classification of internal carotid artery from Doppler signals was presented in [31] and [32]. A number of features were

extracted using PSD and Burg autoregressive spectrum analysis [31]. The Learning Vector Quantization neural network

ACCEPTED MANUSCRIPT

ACCEPTED M
ANUSCRIP

T

was used to classify the signals, and the accuracy rate reported was 97.91%. In [32], using the same feature extraction

technique, a Discrete Hidden Markov Model was used instead of neural networks. The classification accuracy rate was

slightly lower at 97.38% [32]. Carotid artery ultrasound images were processed using information gain based fuzzy c-

means clustering [33]. Various features were then extracted from vectors pertaining to the Intima media thickness values

[33]. Using the PNN, the system performed diagnosis of subjects with an accuracy rate of 98.40% [33].

A method for automatic screening of carotid artery diseases was presented in [34]. Doppler ultrasound signals were first

pre-processed for eliminating noise. Different features from frequency signal domain were extracted with spectrograms

[34]. Using the MLP, the classification rates were 91.67% for normal patterns and 95.89% for occlusion patterns,

respectively [34]. To distinguish benign from malignant, Color Doppler ultrasounds signals were used in [35]. The mean

vector similarity was first used in separating the samples [35]. Three different models were employed for classification, i.e.,

multiple linear regression (MLR), back propagation neural network (BPNN), and genetic algorithm based-BPNN [35]. The

accuracy rates of these three models were 89.5%, 89.5%, and 92.1%, respectively [35]. Table 1 shows a summary of the

literature review presented in this paper.

Table 1 Summary of literature review

Ref Input signals Feature extraction Neural network
Results

Accuracy Specificity Sensitivity

[16] OA Doppler
Least squares

autoregressive
MLP

93.75% to

96.43%

[17] OA and ICA artery
Wavelet transform and

statistical
MLP with LM

96.43% to

97.44%

[18] Doppler ultrasound PSD MLP with LM 94%

[19] Doppler ultrasound STFT MLP with LM 97.8%

[20] Phonocardiogram Multi-resolution wavelet MLP with BP 94.42%

[21] Femoral artery Doppler PSD MLP with LM
98% to

99%

[22] ICA Doppler Wavelet transform MLP with LM
96% to

96.30%

[23] 11 patient variables – MLP 26% 100%

[24] 42 clinical variables –

MLP with Bayesian

regularization

RBF with k-means

86%

88.4%

83.9%

81.8%

[25] Doppler signals PSD
MLP

RBF

94.2%

88.4%

[26] Doppler signals –
MLP

RBF

87.5%

80%

[27] Doppler ultrasound Lyapunov exponents PNN
98.17% to

98.18%

[28] Ophthalmic artery
Resistivity and pulsatility

indices
ANFIS 95.83%

[29]
Mitral valve Doppler

signals
STFT ANFIS 98%

[30] Ulnar artery Doppler MUSIC ANFIS
91.25% to

95%

[31] ICA Doppler PSD LVQ 97.91%

[32] ICA Doppler PSD DHMM 97.38%

[33]
Carotid artery

ultrasound
IMT vectors PNN 98.40%

[34] Doppler ultrasound Spectrogram MLP
91.67% to

95.89%

[35] Color Doppler –

MLR

BPNN

GA-BPNN

89.5%

89.5%

92.1%

3. The recurrent neural network

Among various ANN models, the RNN has the advantage of storing information pertaining to time in a recursive

manner, and is suited for tackling time series related problems [36]. An RNN is similar in structure to that of a standard

feedforward neural network, except that it has feedback links. As shown in Fig. 1, an RNN has connections between the

hidden neurons, which form a directed cycle. Compared with the feedforward neural network, the RNN uses its internal

memory to process an arbitrary sequence of input patterns. While MLP constructs a mapping from the input to output data,

the RNN maps each output to the input’s data sequence. The main difference is the recurrent connections that allow the

input’s historical information in a chain of data sequence to remain in the internal state of the network; therefore influencing

the network output recursively.

ACCEPTED MANUSCRIPT

ACCEPTED M
ANUSCRIP

T

Fig. 1 The recurrent neural network structure (adapted from [37, 38])

3.1 Network dynamics

In an RNN, the forward pass is identical to that of an MLP, except that the activations come to the hidden layer from

both external input and hidden layer information from the previous time step. Consider a length T input sequence x to an

RNN with Q input nodes, N hidden nodes, and G output nodes. Let
as the value of input i at time t,

and
be the

input and activation to node j at time t, respectively [37]. The activation of the n-th hidden node is:

∑

∑

(1)

Similar to an MLP, a nonlinear differentiable activation function is applied, as follows:

(

) (2)

where is the activation function and as the final activation. For a full sequence of the hidden activation, the

calculation is performed using t=1 at the start value, and continuously using eq. (1) and eq. (2) with an increment of t at

each step. One requirement is that the starting value for the hidden node, i.e.,
, needs to be chosen, which corresponds to

the network’s state prior to receiving any information from the data sequence. The starting value is typically set at zero

[37]. However, some studies discovered that network stability and performance could be improved by using non-zero initial

values [39]. The network inputs to the output node can be measured as hidden activations at the same time using:

∑

(3)

The recurrent connections are fixed at 1.0 for connections between the hidden and context layers. The context layer

maintains a copy of the previous values of the hidden nodes. This allows the RNN to keep a state of short-term memory

that facilitates sequence prediction tasks. The other network connections are tuned by collating the targets, and then using

backpropagation to update the connections. As batch training is used, the bias and weights are updated once the targets and

inputs are presented.

3.2 Network learning

For the RNN, a commonly used learning method is the backpropagation algorithm, which adjusts network parameters in

reducing the error measure function with a gradient descent technique. The gradient descent technique aims to change each

weight in proportion to the derivative of the error with respect to the weight [37]. In gradient descent, small fixed-size steps

are repeatedly taken in the negative error gradient direction of the loss function, as follows:

(4)

where Δw
n
indicates the n

th
weight update, L is the loss factor, w

n
indicates the weight vector prior to applying Δw

n
, and α ∈

[0, 1] is the learning rate [37]. The learning iteration is repeated until a stopping criterion is met.

One main issue is that the learning algorithm can be trapped in local minima. This can be avoided by adding a

momentum term [40], which in effect attaches an inertia onto the algorithm motion using the weight space. Using p ∈ [0, 1]

as a momentum parameter, this speeds up the convergence process and allows the learning algorithm to escape from local

minima, as follows:

ACCEPTED MANUSCRIPT

ACCEPTED M
ANUSCRIP

T

(5)

4. Materials and methods

The circle of Willis is a circulatory anastomosis used to supply blood to the human brain and the surrounding structures.

A schematic representation of the circle of Willis, based on the signals used, is shown in Fig. 2. It is divided into two, i.e.,

the anterior and posterior circulations. It comprises the anterior cerebral artery (ACA), anterior communicating artery,

internal carotid artery (ICA), posterior cerebral artery, posterior communicating artery, and basilar artery (BA) [41]. The

middle cerebral artery (MCA) also supplies blood to the brain; however it is not considered as part of the circle [41]. Other

arteries include the ophthalmic artery (OA), intracranial vertebral artery (VA), and carotid siphon (CS).

An experimental study on stenosis detection in collaboration with University Malaya Medical Centre (UMMC) was

conducted. Following the ethics approval from the UMMC Medical Ethics Committee, data from anonymous patients were

retrieved from consecutive TCD studies. Doppler signal acquisition was conducted using a SONARA
TM

TCD unit

(CareFusion, USA) at UMMC, as shown in Fig. 3(a). The probe is located at the right side of the TCD unit. An example of

the acquired Doppler signals is shown in Fig. 3(b).

Fig. 2 A schematic representation of the circle of Willis (adapted from [42])

(a) (b)

Fig. 3 (a) Sonara TCD machine setup (b) acquisition of signals

The patients participated in this study consisted of 14 males and 21 females, who had an established clinical diagnosis of

ischemic stroke. The mean age was 58 years old, with a range from 16 to 86 years old. The patients were classified into

two categories, without intracranial arterial stenosis (class 1) and with stenosis (class 2). In the experimental study, a total

of seven different TCD signals were used, and this represented arterial segments in the circle of Willis i.e. anterior cerebral

arteries (ACA), basilar artery (BA), internal carotid artery (ICA), middle cerebral artery (MCA), ophthalmic artery (OA),

vertebral artery (VA), and carotid siphon (CS). In consultation with a medical expert, a total of five features were extracted

ACCEPTED MANUSCRIPT

ACCEPTED M
ANUSCRIP

T

from each signal, i.e., depth, upper mean flow velocities, lower mean flow velocities, upper pulsatility index, and lower

pulsatility index. The seven different TCD signals with the five extracted features were first fed into the RNN. Based on

individual RNN results, the majority voting method was used to combine the predictions from them. In addition to the

ensemble results, the averages from individual RNN networks were computed. The performance was compared with those

from the RFE model. The experimental procedure is shown in Fig. 4.

Fig. 4 The experimental procedure

5. Results and discussion

The k-fold cross-validation (CV) was used for evaluation, as it could reduce bias linked to random sampling used during

the training stage [43]. With 10-fold CV, the data samples from each TCD signal (i.e. ACA, BA, ICA, MCA, OA, VA, and

CS) were randomly divided into 10 sub-sets, with each having almost an equal number of samples. The RNN was first

trained and then tested for a total of 10 times. In each run, one of the 10 sub-sets was utilized as a test set and the remaining

9 as a training set. The sequence of each training set was randomized. The RNN was configured according to the default

settings in MATLAB. The stopping criterion for the training algorithm was set to a total of 1,000 epochs. The experiment

was repeated five times. In addition, the RFE model [15] was used for performance comparison. The same setup as used

for the RNN was followed, and the best results from a total of 10 trees were selected for each TCD signal.

The performance indicators used in the experiment included accuracy, sensitivity, specificity, and area under the

Receiver Operating Characteristic curve (AUC). To quantify the results statistically, the bootstrap method [44] was

employed. In general, 1,000 bootstrapped samples could give good estimates, while 2,000 bootstrapped samples could give

useful results [44]. In this study, the performance metrics and their respective 95% confidence intervals were computed

using 5,000 bootstrapped samples to provide accurate estimated results.

The overall ensemble and average accuracy, sensitivity, and specificity rates are shown in Fig. 5-7. Note that in-RNN

indicates the average individual RNN results, while en-RNN and

en-RFE

indicate the ensemble results. The error bars

indicate the 95% confidence intervals. The detailed numerical results are presented in Tables 2 to 4, with the 95%

confidence intervals from the 10-fold CV runs. Comparison of the AUC is made as well, as in Table 5.

5.1 Overall results

The accuracy rates are shown in Fig. 5. For individual RNN models, the accuracy rates varied from 70% to 79%. ICA

scored the lowest accuracy rate of 70%, while VA produced the highest accuracy rate of 79%. The results increased using

ensemble RNN, with accuracy rates varying from 74% to 85%. This indicated the usefulness of majority voting in

combining the individual predictions to yield an improved performance. ACA scored the highest accuracy rate of 85%.

The results of RFE ranged from 71% to 82%, with about 3% to 5% lower than those from ensemble RNN. Based on the

error bars, ensemble RNN showed more stable results with smaller 95% confidence intervals. The RFE and individual

RNN results indicated moderate and large variations in terms of the 95% confidence intervals.

The sensitivity rates (Fig. 6) were slightly higher than the specificity rates (Fig. 7) for the experiments. VA from

individual RNN had the highest sensitivity rate of 81%, while ACA from ensemble RNN scored 87% sensitivity. On

average, the sensitivity rates were between 77% and 81% for individual and ensemble RNN models, respectively. On the

other hand, the specificity rates were slightly lower, ranging from 69% to 78% in the individual RNN and 73% to 84% in

the ensemble RNN. VA from individual RNN had the highest specificity rate of 78%, while ACA from ensemble RNN

scored the highest specificity rate of 84%. Comparatively, RFE performed better than individual RNN, but worse than

ensemble RNN. In the data set, there were more samples in class 2 (i.e., with stenosis) as compared with those in class 1

(i.e., without stenosis). This caused more accurate results for class 2 prediction, as compared with those from class 1.

Overall, ACA exhibited the highest accuracy, sensitivity, and specificity rate, as compared to all other signal sources, as

shown in Figs 5 to 7.

Input features
Recurrent

neural network

Outputs Majority voting Final result

ACCEPTED MANUSCRIPT

ACCEPTED M
ANUSCRIP

T

Fig. 5 Accuracy rates from the RNN and RFE

Fig. 6 Sensitivity rates from the RNN and RFE

Fig. 7 Specificity rates from the RNN and RFE

5.2 Detailed results

The detailed numerical scores of accuracy, sensitivity, and specificity are shown in Table 2. In general, individual RNN

produced the lowest rates across all the experiments, while the ensemble RNN scored the highest. Comparing individual

and ensemble RNN models, it could be noticed that the results for BA, MCA, and Siphon overlapped each other in terms of

their 95% confidence intervals. However, the results of ACA, ICA, OA, and VA were distinctive without any overlapping

in their 95% confidence intervals. This indicated that majority voting could improve the individual RNN performance

statistically.

For comparison between ensemble RNN and RFE, none of the 95% confidence intervals pertaining to accuracy,

sensitivity, and specificity overlapped one another. The 95% confidence intervals of RFE were slightly larger than those of

ensemble RNN. On average, up to 4% difference in the average accuracy score could be noticed between the two models.

On the other hand, RFE performed similarly with individual RNN, owing to the overlap of their large 95% confidence

intervals. In general, ACA yielded the best results from ensemble RNN and RFE, while the second best from individual

RNN.

Table 2 Accuracy rates from individual and ensemble RNN models and RFE

Individual model Ensemble model

Signal RNN RNN RFE

BA 73.33% (69.19%–77.47%) 73.84% (72.76%–74.92%) 71.89% (69.77%–72.00%)

ACA 78.54% (74.41%–82.67%) 85.02% (84.01%–86.03%) 81.77% (79.73%–81.8%)

ICA 70.00% (67.16%–72.84%) 75.16% (74.40%–75.92%) 72.15% (71.37%–72.93%)

MCA 73.53% (69.53%–77.53%) 75.89% (75.00%–76.78%) 71.1% (71.18%–73.01%)

OA 75.06% (71.75%–78.37%) 80.97% (80.12%–81.82%) 76.73% (76.85%–78.61%)

Siphon 77.49% (73.91%–81.07%) 80.88% (80.02%–81.74%) 77.84% (75.95%–77.72%)

VA 78.80% (75.80%–81.80%) 83.16% (82.37%–83.95%) 80.83% (79.02%–80.65%)

60.00%

70.00%

80.00%

90.00%

BA ACA ICA MCA OA Siphon VA

in-RNN

en-RNN

en-RFE

60.00%

70.00%

80.00%

90.00%

BA ACA ICA MCA OA Siphon VA

in-RNN

en-RNN

en-RFE

60.00%

70.00%

80.00%

90.00%

BA ACA ICA MCA OA Siphon VA

in-RNN

en-RNN

en-RFE

Naima_Nisha
Highlight

ACCEPTED MANUSCRIPT

ACCEPTED M
ANUSCRIP

T

Table 3 Sensitivity rates from individual and ensemble RNN models and RFE

Individual model Ensemble model

Signal RNN RNN RFE

BA 75.33% (71.53%–79.13%) 75.84% (74.40%–77.28%) 72.81% (71.32%–74.29%)

ACA 80.54% (77.14%–83.94%) 87.02% (85.69%–88.35%) 83.67% (81.3%–84.04%)

ICA 72.00% (69.29%–74.71%) 77.16% (76.15%–78.17%) 74.07% (73.03%–75.12%)

MCA 75.53% (71.61%–79.45%) 77.89% (76.72%–79.06%) 74.60% (72.79%–75.20%)

OA 77.06% (73.85%–80.27%) 82.97% (81.85%–84.09%) 79.65% (78.50%–80.80%)

Siphon 79.49% (75.60%–83.38%) 82.88% (81.72%–84.04%) 78.94% (77.54%–79.93%)

VA 80.80% (78.06%–83.54%) 85.16% (84.11%–86.21%) 82.25% (80.67%–82.83%)

Table 4 Specificity rates from individual and ensemble RNN models and RFE

Individual model Ensemble model

Signal RNN RNN RFE

BA 72.33% (67.99%–76.67%) 72.84% (71.42%–74.28%) 69.93% (68.44%–71.41%)

ACA 77.54% (73.59%–81.49%) 84.02% (82.63%–85.41%) 80.82% (78.38%–81.25%)

ICA 69.00% (66.02%–71.98%) 74.16% (73.14%–75.18%) 72.19% (70.14%–72.24%)

MCA 72.53% (69.13%–75.93%) 74.89% (73.72%–76.06%) 71.15% (69.94%–72.35%)

OA 74.06% (70.78%–77.34%) 79.97% (78.85%–81.09%) 76.77% (75.62%–77.93%)

Siphon 76.49% (72.71%–80.27%) 79.88% (78.75%–81.01%) 76.89% (74.73%–77.05%)

VA 77.80% (75.04%–80.56%) 82.16% (81.12%–83.20%) 79.87% (77.80%–79.95%)

5.3 AUC results

Another important performance metric in medical applications, i.e., the AUC, was computed for performance

comparison, as shown in Table 5. Overall, the AUC scores varied between 0.71 and 0.86. The highest AUC was produced

by ensemble RNN using signals from ACA (i.e., 0.8552), while the lowest was by individual RNN with ICA (i.e., 0.7050).

RFE outperformed individual RNN in terms of ACA, ICA, OA, while the opposite could be observed for BA, MCA, and

Siphon. Ensemble RNN yielded the best AUC scores for all signal sources; therefore indicating its effectiveness in

diagnosis stenosis using TCD signals.

Table 5 AUC scores from individual and ensemble RNN models and RFE

Individual model Ensemble model

Signal RNN RNN RFE

BA 0.7383 0.7434 0.7089

ACA 0.7904 0.8552 0.8077

ICA 0.7050 0.7566 0.7215

MCA 0.7403 0.7639 0.7210

OA 0.7556 0.8147 0.7773

Siphon 0.7799 0.8138 0.7684

VA 0.7930 0.8366 0.7983

6. Conclusions

In this paper, an analysis of TCD signals recorded from the arterial segments of the circle of Willis from real patients

with ischemic stroke has been examined. TCD signals from a total of 35 anonymous patients have been acquired for

evaluation using individual and ensemble RNN models as well as RFE. Ensemble RNN has shown good performances with

mean accuracy higher than 70%, with the highest accuracy rate at 85%. In addition to RNN, RFE has been used for

performance comparison. Comparing with individual RNN and RFE, it is evident that ensemble RNN is able to make

accurate predictions using the majority voting method. The AUC results from ensemble RNN also indicate good

performance, with the highest AUC score of 0.8552.

For further work, an online classifier for detection of TCD signals, which can be used in a hospital environment, will be

developed. In addition, other medical prognostic and diagnostic applications of ensemble RNN will be investigation. The

ultimate goal is to promote the use of machine learning models to help medical practitioners in detecting medical

abnormalities and making a rapid and informed decision to treat patients.

Naima_Nisha
Highlight

Naima_Nisha
Highlight

ACCEPTED MANUSCRIPT

ACCEPTED M
ANUSCRIP

T

References

[1] Ozturk, A., & Arslan, A. (2007). Classification of transcranial Doppler signals using their chaotic invariant measures. Computer

Methods and Programs in Biomedicine, 86(2), 171-180.

[2] Uğuz, H. (2012). A hybrid system based on information gain and principal component analysis for the classification of transcranial

Doppler signals. Computer Methods and Programs in Biomedicine, 107(3), 598-609.

[3] Austen, W. G., & Howry, D. H. (1965). Ultrasound as a method to detect bubbles or particulate matter in the arterial line during

cardiopulmonary bypass. Journal of Surgical Research, 5(6), 283-284.

[4] Azarpazhooh, M. R., & Chambers, B. R. (2006). Clinical application of transcranial Doppler monitoring for embolic signals. Journal

of clinical neuroscience, 13(8), 799-810.

[5] Karaboga, N., & Latifoglu, F. (2013). Adaptive filtering noisy transcranial Doppler signal by using artificial bee colony algorithm.

Engineering Applications of Artificial Intelligence, 26(2), 677-684.

[6] Özbay, Y., Kara, S., Latifoğlu, F., Ceylan, R., & Ceylan, M. (2007). Complex-valued wavelet artificial neural network for Doppler

signals classifying. Artificial intelligence in medicine, 40(2), 143-156.

[7] Aaslid, R., Markwalder, T. M., & Nornes, H. (1982). Noninvasive transcranial Doppler ultrasound recording of flow velocity in

basal cerebral arteries. Journal of neurosurgery, 57(6), 769-774.

[8] Alam, J., Hassan, M., Khan, A., & Chaudhry, A. (2015). Robust fuzzy RBF network based image segmentation and intelligent

decision making system for carotid artery ultrasound images. Neurocomputing, 151, 745-755.

[9] Kasabov, N., Feigin, V., Hou, Z. G., Chen, Y., Liang, L., Krishnamurthi, R., & Parmar, P. (2014). Evolving spiking neural networks

for personalised modelling, classification and prediction of spatio-temporal patterns with a case study on

stroke. Neurocomputing, 134, 269-279.

[10] Walter, A., Naros, G., Spüler, M., Gharabaghi, A., Rosenstiel, W., & Bogdan, M. (2014). Decoding stimulation intensity from

evoked ECoG activity.Neurocomputing, 141, 46-53.

[11] Ceylan, M., Ceylan, R., Dirgenali, F., Kara, S., & Özbay, Y. (2007). Classification of carotid artery Doppler signals in the early

phase of atherosclerosis using complex-valued artificial neural network. Computers in Biology and Medicine, 37(1), 28-36.

[12] Özbay, Y., & Ceylan, M. (2007). Effects of window types on classification of carotid artery Doppler signals in the early phase of

atherosclerosis using complex-valued artificial neural network. Computers in Biology and Medicine, 37(3), 287-295.

[13] Ceylan, M., Ceylan, R., Özbay, Y., & Kara, S. (2008). Application of complex discrete wavelet transform in classification of

Doppler signals using complex-valued artificial neural network. Artificial intelligence in medicine, 44(1), 65-76.

[14] Shieh, J. S., Chou, C. F., Huang, S. J., & Kao, M. C. (2004). Intracranial pressure model in intensive care unit using a simple

recurrent neural network through time. Neurocomputing, 57, 239-256.

[15] Han J, Kamber M, Pei J (2012). Data Mining: Concepts and Techniques (3rd ed.), MA, USA: Morgan Kaufmann.

[16] Güler, İ., & Übeyli, E. D. (2005). Detection of ophthalmic arterial doppler signals with Behcet disease using multilayer perceptron

neural network. Computers in biology and medicine, 35(2), 121-132.

[17] Übeyli, E. D., & Güler, İ. (2005). Improving medical diagnostic accuracy of ultrasound Doppler signals by combining neural

network models. Computers in Biology and Medicine, 35(6), 533-554.

[18] Kara, S., Güven, A., Okandan, M., & Dirgenali, F. (2006). Utilization of artificial neural networks and autoregressive modeling in

diagnosing mitral valve stenosis. Computers in biology and medicine, 36(5), 473-483.

[19] Kara, S. (2007). Classification of mitral stenosis from Doppler signals using short time Fourier transform and artificial neural

networks. Expert Systems with Applications, 33(2), 468-475.

[20] Babaei, S., & Geranmayeh, A. (2009). Heart sound reproduction based on neural network classification of cardiac valve disorders

using wavelet transforms of PCG signals. Computers in Biology and Medicine, 39(1), 8-15.

[21] Kara, S., Kemaloğlu, S., & Güven, A. (2005). Detection of femoral artery occlusion from spectral density of Doppler signals using

the artificial neural network. Expert Systems with Applications, 29(4), 945-952.

[22] Übeyli, E. D., & Güler, İ. (2006). Wavelet-based neural network analysis of internal carotid arterial Doppler signals. Journal of

medical systems, 30(3), 221-229.

[23] Mobley, B. A., Schechter, E., Moore, W. E., McKee, P. A., & Eichner, J. E. (2005). Neural network predictions of significant

coronary artery stenosis in men. Artificial intelligence in medicine, 34(2), 151-161.

[24] Mehrabi, S., Maghsoudloo, M., Arabalibeik, H., Noormand, R., & Nozari, Y. (2009). Application of multilayer perceptron and radial

basis function neural networks in differentiating between chronic obstructive pulmonary and congestive heart failure diseases. Expert

Systems with Applications, 36(3), 6956-6959.

[25] Ergün, U., Barýþçý, N., Ozan, A. T., Serhatlýoðlu, S., Oğur, E., Hardalaç, F., & Güler, İ. (2004). Classification of MCA stenosis in

diabetes by MLP and RBF neural network. Journal of Medical Systems, 28(5), 475-487.

[26] Yýldýrým, H., Altýnsoy, H. B., Barýpçý, N., Ergün, U., Oğur, E., Hardalaç, F., & Güler, İ. (2004). Classification of the frequency of

carotid artery stenosis with MLP and RBF neural networks in patients with coroner artery disease. Journal of medical systems, 28(6),

591-601.

[27] Übeyli, E. D. (2008). Probabilistic neural networks employing Lyapunov exponents for analysis of Doppler ultrasound signals.

Computers in biology and medicine, 38(1), 82-89.

[28] Güler, İ., & Übeyli, E. D. (2005). Automatic detection of ophthalmic artery stenosis using the adaptive neuro-fuzzy inference

system. Engineering Applications of Artificial Intelligence, 18(4), 413-422.

[29] Tokmakçi, M. (2007). A classification system for stenosis from mitral valve Doppler signals using adaptive network based fuzzy

inference system. Journal of medical systems, 31(5), 329-336.

[30] Özkan, A. O., Kara, S., Salli, A., Sakarya, M. E., & Güneş, S. (2010). Medical diagnosis of rheumatoid arthritis disease from right

and left hand Ulnar artery Doppler signals using adaptive network based fuzzy inference system (ANFIS) and MUSIC

method. Advances in

[31] Uğuz, H. (2012). Detection of carotid artery disease by using learning vector quantization neural network. Journal of medical

systems, 36(2), 533-540.

[32] Uğuz, H., & Kodaz, H. (2011). Classification of internal carotid artery Doppler signals using fuzzy discrete hidden Markov

model. Expert Systems with Applications, 38(6), 7407-7414.

[33] Hassan, M., Chaudhry, A., Khan, A., & Iftikhar, M. A. (2014). Robust information gain based fuzzy c-means clustering and

ACCEPTED MANUSCRIPT

ACCEPTED M
ANUSCRIP

T

classification of carotid artery ultrasound images. Computer methods and programs in biomedicine,113(2), 593-609.

[34] Seddik, A. F., & Shawky, D. M. (2013). A low-cost screening method for the detection of the carotid artery diseases. Knowledge-

Based Systems, 52, 236-245.

[35] Chen, X. L., Yang, L. M., Wang, S. C., Yang, R., & Wang, J. X. (2011). Identifying the nature of stomach diseases by

ultrasonography based on genetic neural network. Expert Systems with Applications, 38(5), 6400-6403.

[36] Rowe, R. & Roberts, R. J. (2011). Intelligent software for product formulation. Taylor & Francis, London.

[37] Graves, A. (2012). Supervised sequence labelling with recurrent neural networks (Vol. 385). Heidelberg: Springer.

[38] Jain, L. C., Seera, M., Lim, C. P., & Balasubramaniam, P. (2014). A review of online learning in supervised neural networks. Neural

Computing and Applications, 25(3-4), 491-509.

[39] Zimmermann, H. G., Grothmann, R., Schaefer, A. M., & Tietz, C. (2006). Identification and forecasting of large dynamical systems

by dynamical consistent neural networks. New Directions in Statistical Signal Processing: From Systems to Brain, 203-242.

[40] Plaut, D. C., Nowlan, S. J., & Hinton, G. E. (1986). Experiments on Learning by Back-Propagation. Technical Report CMU–CS–

86–126, Carnegie–Mellon University.

[41] Purves, D., Augustine, G. J., Fitzpatrick, D., Hall, W. C., LaMantia, A.-S., McNamara, J. O., & White, L. E. (2008). Neuroscience,

4th Ed., 834–835.

[42] Circle of Willis (2015). In Wikipedia, Retrieved October 1, 2015, from https://en.wikipedia.org/wiki/Circle_of_Willis

[43] Delen, D., Walker, G., & Kadam, A. (2005). Predicting breast cancer survivability: a comparison of three data mining methods.

Artificial Intelligence in Medicine, 34(2), 113-127.

[44] Efron, B. & Tibshirani, R. (1993). An Introduction to the Bootstrap, Chapman and Hall, New York.