Вы находитесь на странице: 1из 5

A Different approach for Cardiac Arrhythmia Classification using Random Forest

Algorithm
Allam Jaya prakash,Samit Ari
Department of Electronics and Communication Engineering, National Institute of Technology, Rourkela,India
E-mail: allamjayaprakash@gmail.com

Published in Healthcare Technology Letters; Received on xxx; Revised on xxx.

ECG plays a most important role in finding cardiac disorders of the heart. Cardiac arrhythmias occur in a short duration of time which can’t
be distinguishable by a human eye. The finding of arrhythmias is a tedious task since slight changes in the electrocardiogram signal (ECG)
may lead to life-threatening. Diagnosis and medication at an early stage of cardiac arrhythmia may facilitate to decrease the mortality rate of
the heart patients. This work presents an accurate system for the classification of five types of ECG arrhythmias namely Paced (P), Premature
ventricular contraction (V), Normal (N), Right bundle branch block (R) and Left bundle branch block (L). In the proposed technique, to
classify ECG arrhythmias random forest (RF) classifier is applied. The projected model of cardiac arrhythmia recognition system covers 3
stages they’re pre-processing, feature extraction & classification. In the initial stage, filtered the ECG signal raw data and finding the R-peak
locations of the ECG signal. Dual tree complex wavelet transform (DTCWT) is employed to extract the feature vector from the ECG signal
within the second stage. The final feature vector consists of an extracted feature set using DTCWT and four other temporal features of the
ECG signal. The random forest (RF) classifier access final feature set as input. Classifier performance is evaluated using MIT-BIH physionet
database. Finally, RF classifier achieved overall accuracy of 98.78% on an individual basis once tested over five kinds of physionet standard
arrhythmia database (MIT-BIH). The proposed arrangement of framework effectively classified five kinds of ECG arrhythmias. The proposed
RF classifier performs better than other reported techniques for the ECG arrhythmia classification.

1. Introduction: Electrocardiogram (ECG) is the recording of customized ECG beats. Linear vector quantization (LVQ) and self-
the electrical phenomenon at the heart that shows the variability of organizing maps (SOM) are two classification models based on
heart rate. Any deviation from the normal heart rate, disturbance clustering technique, LVQ is a supervised learning approach which
in rhythm, site of origin or conduction of heart electric impulse is classifies the feature vector corresponding to the class label of
considered as an arrhythmia. Most of the arrhythmias are seldom the cluster pattern (code word) into which is clustered. In SOM,
occurring and can’t be distinguished by a person’s eye. Long each centring of the cluster (prototype or code word) is denoted
duration of ECG record called as ambulatory electrocardiogram are by some weights of a neuron which is assigned to correlate in
required to identify the abnormalities present in a patient. Moreover, the feature map. The classifier is modelled using mixture of SOM
huge variations in temporal and some other morphological and LVQ approaches. The network is designed in such a way that
characteristics from one patient to the other patient make detection LVQ gives superior classification performance for classes 1 and 3
of abnormalities is a challenging task. Hence it is very difficult however for class 2 and 4 SOM gives superior performance. Overall
to analyse and recognize these arrhythmias manually by an expert classification accuracy of 94% is reported using this mixture of
cardiologist also. Hence we require an automatic computer-aided expert’s approach.
diagnostic (CAD) system that can quickly detect abnormalities A unique technique is proposed in [7] for a patient-adapted ECG
within the patient. heartbeat classification that consists of four stages namely, pre-
In literature, there are many algorithms proposed on automatic processing, feature extraction, feature selection and classification.
classification of ECG arrhythmias. Most of them have done ECG Features are extracted in temporal and morphological domain.
classification in the subsequent steps i) pre-processing ii) feature morphological domain features include coefficients obtained by
extraction iii) classification. In this feature extraction techniques applying stock well transform. 184 samples around each R peak
include both temporal (time) and morphological domain features. are taken by combining 4 temporal features (Pre R-R, Post
Some of the temporal features include R-R intervals, Q-R intervals, R-R, average R-R, local R-R) and 180 samples of frequency
QRS complex duration, R-S intervals, ST time segments. These domain features and obtained better classification results but poor
temporal features are not enough for classification since there is sensitivity as reported in [7].
a huge variation in the other morphological patterns within the To address the above limitations, a novel approach of random
same patient. Hence some of the researchers have reported the forest (RF) classifier is proposed to classify ECG arrhythmias.
mixture of both frequency and time features as reported in [1]. Random forest is a highly powerful machine learning algorithm
Many transformation techniques are used, like S transform, Fourier which is based on supervised learning.[8] Random forest algorithm
transform (FT) and discrete wavelet transform (DWT) to extract can use individually for both classification and the regression kind
the features from the pre-processed data [2]. Extracted mixture of difficulties. As name suggests that this algorithm creates the
feature set is given to the input of a classifier for classification forest with many number of random trees .
into respective classes. Some of the classifiers are artificial neural The remaining paper is organized as follows: Section 2 presents
networks (ANN) [3], support vector machine (SVM) [4] etc. MIT-BIH database that is used to evaluate the performance of
Detection of cardiac arrhythmias by applying Hidden Markov the proposed classification algorithm. Section 3 contains proposed
models is reported in [5].This algorithm has demonstrated very framework. Section 4 represents the experiment results and
good promising results in classifying ventricular arrhythmias and discussion of the proposed algorithm. The conclusion of the work
detection of low amplitude P wave detection. In [6], the author is explained in section 5.
presented a mixture of experts (MOE) methodology to classify

Healthcare Technology Letters, pp. 1–5 1


The Institution of Engineering and Technology 2018
c
2. ECG data Processing: MIT-BIH arrhythmia database is
ECG Data Acquisition
utilised for assessing the performance of the RF classifier technique.
The MIT-BIH arrhythmia database consists of some deadly
arrhythmias. Most of the researchers consider as the standard

(i) Pre-processing
Normalization
database for detection of cardiac arrhythmias. The database
contains forty-eight files of ECG recordings, and every file consists
half an hour ECG segment selected from 24 hours recordings of Filtering
48 specific patients. The first 23 recordings related to the normal
routine clinical recordings while the remaining recordings contain R-peak detection
the complex ventricular, junctional & supraventricular arrhythmias
[9]. MIT-BIH ECG recordings are digitized at 360 Hz and band-
pass filtered at 0.1-100 Hz. Fig.1 represents the plot of record 100
from MIT-BIH ECG signal database. Annotation file contains labels

Morphological features
Temporal features
AC power
for each rhythm that are detected by using a simple slope sensitive

(ii) Feature Extraction


Select a window ( 192
detector. These labels are utilized in the training and testing phase of samples from right and
Kurtosis
the implemented classifier algorithm. In this work, 5 different types left) around the R-peak
of cardiac arrhythmias namely paced (P), premature ventricular Skewness
contraction (V), Normal (N), right bundle branch block (RBBB)
& left bundle branch block (LBBB) arrhythmias are classified by Timing
DTCWT features
implementing RF algorithm. information

Extracted Features Extracted Features

Total Feature Vector

Random Forest Classifier

P V N RBBB LBBB

Figure 2: Block diagram of proposed methodology

from QRS complex of respective cardiac cycle. If r[n] represents


Figure 1: Plot of 100m ECG signal record from MIT-BIH database the QRS complex signal.
(1) AC power: It represents the total power content in
Electrocardiogram QRS complex signal.
3. Proposed Framework: The complete proposed methodology
for the cardiac arrhythmia classification is shown in Fig.2 which p =E(r[n]2 ) (1)
includes the acquisition of ECG data, pre-processing, feature (2) Kurtosis: It indicates the sharpedness of the Electrocardiogram
extraction and classification. QRS complex signal.

3.1. Pre-processing: The preprocessing stage contains the E[(r − µ)4 ]


kurt(r) = (2)
subsequent two steps: (i) ECG signal amplitudes are normalised σ4
to a zero mean, and therefore the amplitude variance for every
(3) Skewness: More precisely skewness indicates the shortage of
ECG signal is terminated. (ii) Every ECG signal is filtered
symmetry of the distribution or data set.
using a bandpass filter at 0.1-100 Hz to get rid of the noises.
R-peak detection is also included in this stage. The R-peak E[(r − µ)3 ]
of the electrocardiogram signal is determined by the famous skew(r) = (3)
σ3
Pan-Tompkins algorithm [10] from the pre-processed ECG signal.
(4) Timing information: It is the measure of deviation from constant
beat rate calculated by using R-R interval ratio.
3.2. Feature extraction: Extracting unique good features provide
better classification accuracy. In this work we extracted four Tj − Tj−1
temporal features AC power, kurtosis, skewness, timing IRj = (4)
Tj+1 − Tj
information from ECG signal but these features alone are not
sufficient to classify the arrhythmias. Therefore these features are where Tj indicates the time at which meant for beat J occurs.
appended to the morphological features extracted with the help
of Dual tree complex wavelet transform (DTCWT) [11]. Feature 3.2.2. Morphological features: These features are extracted with
extracted portion is described in the Fig.2 block diagram. the help of DTCWT [11]. The physionet database ECG signals are
sampled at 360 samples/sec so a consequence frequency component
3.2.1. Temporal features: Temporal features are also called as time in ECG signal is in the range of 0-180 Hz. In this work coefficients
domain features and these are very easy to extract from the signal. of the DTCWT calculated across the QRS complexes of the
AC power, kurtosis, skewness, timing information are extracted electrocardiogram signal.

2 Healthcare Technology Letters, pp. 2–5



c The Institution of Engineering and Technology 2018
DTCWT is a powerful technique and it is not sufferred with shift classification model. Each random tree established a classification
variance, oscillation and aliasing like DWT and other techniques. based on its trained model and vote for one arrhythmia class among
Due to these abilities DTCWT is preferrable for extracting all the five classes (P,V,N,R,L). The Random forest algorithm is
morphological features of the QRS complexes. The detection and an approach with several decision trees [14]. In this work 10-fold
reconstruction of a signal having singularities become very difficult cross validation is used to avoid overfitting problems. Training set
due to the oscillation of the coefficients between +ve and -ve consists of 90% of the whole data and the remaining 10% were used
values around the singularities in DWT. This problem overcome to test the model [15].
by DTCWT.
4. Experimental results:

4.1. Data selection: In this piece of work, Performance of the


proposed RF classifier is evaluated by using 48 files of MIT-BIH
arrhythmia database and each file accomodate 30 mins of ECG
signal portion choosen from the 24 hours recording of 48 different
patient files. All the MIT-BIH signals are sampled at 360 hertz.The
following five types of arrhythmias named as Paced (P), Premature
ventricular contraction (V), Normal (N), Right bundle branch block
(R) and Left bundle branch block (L) are classified.

4.2. Metrics: In this paper, we compared proposed method to four


recent popular methods used for arrhythmia classification interms
of accuracy (Acc), Sensitivity (Se), Specificity (Sp) [16].
Accuracy is defined as the ratio of perfectly classified patterns to
Figure 3: Three levels of DTCWT the overall patterns applied.
TP + TN
Dual tree CWT has complex valued wavelet and scaling function Accuracy(Acc) = (7)
TP + TN + FP + FN
described below
Sensitivity is also called true positive rate of the classifer.
ϕc (t) = ϕreal (t) + jϕim (t) (5)
TP
Sensitivity(Sen) = = (8)
φc (t) = φreal (t) + jφim (t) (6) TP + FN
The dual tree complex wavelet uses 2 real wavelet filters: one is for Specificity is defined as the rate of correctly classified non events
collecting real part and another one is for collecting imaginary part among all non events.
of the resulting transform. Fig.3 shows 3 levels of DTCWT. Tree A TP + TN
handles the real part and tree B handles the imaginary part of the of Specificity(Spe) = (9)
TN + FP
the complex wavelet transform.
Positive-Predictivity is defined as the rate of correctly classified
events in all recognized events.
3.3. Classifier: Random forest (RF) is a smart tool for multi-
classification and regression problems with good classification TP
Positive_Predictivity(PPr) = (10)
accuracy. In general random forest classifier is the extension TP + FP
of random tree classifier and it has multiple decision trees.[12]
where TP, TN, FP, FN are true positive, true negative, false positive,
RF classifier was invented by Leo breiman from university of
false negative respectively which are calculated from the confusion
California,United states. RF classifier has the capability to handle
matrix.
large number of input data. RF algorithm integrates multiple
Table 1 represents relative performance of proposed work with
decision trees and provides overall aggregate predictions of those
the recent works on ECG arrhythmia classification based on
trees.
physionet database.The disadvantage of Chazal et al. technique is
that fixed classification technique doesn’t consider any variation
in ECG pattern caused by personal or environmental differences.
Chazal et al. are used AAMI standard five heart beat types
and only they used 44 files for the experimental study with
classification accuracy 85.90%. In [17] the authors used lyapunov
exponents,wavelet coefficients and they consider power levels of
power spectral density (PSD) values of the Electrocardiogram
signal as features for classifying four types of ECG beats of
MIT-BIH database. In the reported work [6] experimental study
conducted on very small data set and DWT is used to extract the
features, which lacks the property of shift invariance due to the
down sampling operation at each stage of DWT implementation.
Due to these reasons classification accuracy achieved only 93.90%.
The diagonal elements of the confusion matrix specifies the
Figure 4: Generalised random forest structure
perfectly classified instances corresponding to their individual
classes. From Table 2, it is clear that 197 N arrhythmia beats,
In the proposed random forest applies decision random trees to 398 V arrhythmia beats, 67 P arrhythmia beats, 310 L arrhythmia
classify arrhythmias based on applied features, where the applied beats, 310 R arrhythmia beats are misclassified. Fig.5 represents
features are extracted by using DTCWT technique [13].During predicted patterns of five types of arrhythmias.V arrhythmia beat is
random forest training, individual decision tree utilise a subgroup very important in clinical diagnosis for cardiac arrhythmia patient.
of the features (along with labels information) to develop output Compared to the previous techniques random forest classified V

Healthcare Technology Letters, pp. 3–5 3


The Institution of Engineering and Technology 2018
c
Table 1 Summary of the techniques on ECG arrhythmia classification using MIT-BIH database
Literature Features Classifier Classes Accuracy (%)
De Chazal et al. Morphology and heart beat interval Linear discriminant 5 85.90
Ubeyli Lyapunov exponents and wavelet coefficients ANN classifier 4 93.90
Hu et al. Time domain features Mixture of experts 2 94.00
Manu Thomas et al. DTCWT features ANN classifier 5 94.64
Proposed method DTCWT features Random forest 5 98.78

beat with better accuracy.The proposed method classifies N, P, L, R arrhythmia beat detection shows an accuracy of 99.70%,
R, V with sensitivity of 99.74%, 99.05%, 96.16%, 95.88%, 94.42% sensitivity of 95.88%, specificity of 99.99% and positive
respectively. The proposed technique provides best overall accuracy predictivity of 99.86%. F-score is calculated based on precision
of 98.78% compared to the all other classifiers. The proposed and recall. F-score provides the most realistic measure of classifier
technique with random classifier gives higher true positive rate performance.
(TPR) than false positive rate (FPR) [18].
In [11] M.Thomas et al. selected a window of 128 samples
from right and left of the R-peak of the ECG signal.The final
feature vector consists of an extracted feature set from DTCWT and
four other temporal features.The authors proposed multilayer back
propagation neural network is proposed to classify the five types
of ECG arrhythmias. Multilayer back propogation neural network
is very much sensitive to the hidden layer neurons. When hidden
layer neurons are less,then MLP leads to underfitting. If the hidden
layer neurons are many then it leads to over fitting so that the
fitting curve takes uncontrolled oscillations.Network dysfunction
happens once the weights are adjusted to terribly massive values.
It has the following disadvantages (i) A fixed handcrafted feature
extraction method may not be suitable for extracting features when
classifying beats of a particular person (ii) Complexity in the
structure of this technique imposes difficulty in implementation.
DTCWT is a promising technique even though because of multi Figure 5: Patterns classification plot of N, V, P, L, R arrhythmias
layer back propogation drawbacks final classification accuracy
decreased. These drawbacks can be overcome by random forest
Precision and recall values are calculated using confusion matrix
algorithm to get better accuracy [19].
represented in Table 2. Precision is defined as the fraction of
Table 3 represents the performance parameters of the classifier
instances correctly classified as positive out of all the instances
like accuracy (Acc), sensitivity (Sen), specificity (Spe), positive-
the algorithm classified as positive, whereas recall is the fraction
predictivity (Ppr) [20]. All 48 files including four files of paced
of elements correctly classified as positive out of all the positive
beats also included for the performance evaluation of the classifier.
elements [21].
For 48 files the N arrhythmia beat detection shows an accuracy of
98.93% , sensitivity of 99.74%, specificity of 96.85% and positive 2 ∗ TP
predictivity of 98.78%, V arrhythmia beat detection shows an F − score(F) = (11)
2 ∗ TP + FP + FN
accuracy of 99.39% , sensitivity of 94.42%, specificity of 99.76%
and positive predictivity of 96.63%, P arrhythmia beat detection The overall F-score of a classifier calculated as 98.24 by using
shows an accuracy of 99.90% , sensitivity of 99.05%, specificity precision and recall values.
of 99.96% and positive predictivity of 99.51%. L arrhythmia beat
detection shows an accuracy of 99.63% , sensitivity of 96.16%, Table 4 True, False positive rates & Roc area of Individual Classes
specificity of 99.93% and positive predictivity of 99.11%. Class TP Rate FP Rate ROC Area
N 0.997 0.031 0.999
Table 2 Classification results of MIT-BIH arrhythmia database V 0.944 0.002 0.999
P 0.990 0.000 1.000
Class Confusion matrix [DTCWT] L 0.962 0.001 0.999
N V P L R R 0.959 0.000 0.999
N 74788 142 9 42 4
V 357 6730 21 17 3
P 43 23 6953 1 0 ROC area is used to evaluate the performance of the classifier
L 250 54 3 7759 3 [22]. An area of 1.000, 0.500 represents the perfect test and test
R 272 16 1 10 6952 fail respectively. Table 4 shows the ROC area of individual five
arrhythmia classes. For all classes ROC area is above 0.9. The
results shows that Random forest classifier works efficiently for the
Table 3 Classification performance of RF with DTCWT features
arrhythmia classification [23].
Performance matrix
Method Class
Acc (%) Sen (%) Spe (%) Ppr (%) 5. Conclusion: In this letter, an automatic classification
N 98.93 99.74 96.85 98.78 approach is proposed to classify the ECG arrhythmia types of
V 99.39 94.42 99.76 96.63 Paced (P), Premature ventricular contraction (V), Normal (N),
Proposed
P 99.90 99.05 99.96 99.51 Right bundle branch block (R) and Left bundle branch block
Method
L 99.63 96.16 99.93 99.11 (L). The experiments are conducted on the standard MIT-BIH
R 99.70 95.88 99.99 99.86 standard physionet dataset [24]. In this work Feature extraction

4 Healthcare Technology Letters, pp. 4–5



c The Institution of Engineering and Technology 2018
and classification are the two important steps to detect the [14] J. J. Rodriguez., L. I. Kuncheva and C. J. Alonso.:
ECG arrhythmia type. In the proposed method DTCWT extract ‘Rotation Forest: A New Classifier Ensemble Method’,
the morphological features of the ECG signal and additionally IEEE Transactions on Pattern Analysis and Machine
four other temporal features (AC power, Kurtosis, Skewness and Intelligence., 2006, 28,(10), pp. 1619-1630
Timing Information) are also extracted from the ECG signal [25]. [15] T.K. Ho.: ‘A Data Complexity Analysis of Comparative
Morphological features are appended with temporal features and Advantages of Decision Forest Constructors’, Pattern
considered as final feature vector. This Final feature vector used Analysis and Applications., 2002, 5,(4), pp. 102-112
as an input to the random forest classifier. Table 3 suggest that [16] R. C. Prati., G. E. A. P. A. Batista and M. C. Monard.:
the chosen feature extraction method and classifier provides high ‘Evaluating Classifiers Using ROC Curves’, IEEE Latin
overall accuracy of 98.78%, sensitivity of 98.24% , specificity of
America Transactions., 2008, 6,(2), pp. 215-222
98.12% and positive predictivity of 97.89%. These results show
[17] Ubeyli ED.: ‘Statistics over features of ECG signals’, Expert
that an advance improvement is attained for the proposed method
Syst Appl., 2009, 36, (5), pp. 8758-67
of ECG arrhythmia classification when compared to other recent
[18] M. F. Amasyali and O. K. Ersoy.: ‘Rotation Forest:
methods.
Classifier Ensembles with the Extended Space Forest’, IEEE
Transactions on Knowledge and Data Engineering., 2014,
6 References 26,(03), pp. 549-562
[19] Y. Zhou and G. Qiu.: ‘Random forest for label ranking’,
[1] Inan O.T., L. Giovangrandi., G.T.A Kovacs.: ‘Robust Expert Systems with Applications., 2018, 112,(06), pp. 99-
neural network based classification of premature ventricular 109
contractions using wavelet transform and timing interval [20] M. Zabihi., A. B. Rad.,A. K. Katsaggelos.,S. Kiranyaz., S.
features’, IEEE trans. Inf.Technol. Biomed., 2006, 53, (12), Narkilahti and M. Gabbouj.: ‘Detection of atrial fibrillation
pp. 2507-2515 in ECG hand-held devices using a random forest classifier’,
[2] D. Cvetkovic., E. D. U beyli., I. Cosic.: ‘Wavelet transform Computing in Cardiology (CinC)., 2017, pp. 1-4
feature extraction from human ppg , ecg, and eeg signal [21] Liamedo M., Martinez J.P.: ‘An automatic patient-adapted
responses to elf pemf exposures: A pilot study’, Digital ECG heartbeat classifier allowing expert assistance’, IEEE
signal processing., 2008, 18, (5), pp. 861-874 Trans. Biomed. Eng., 2012, 59, (8), pp. 2312-2320
[3] J. J. Oresko., Z. Jin., J. Cheng., S. Huang., Y. Sun., H. [22] P. De Chazal., M. O. Dwyer.,R. B. Reilly.: ‘Automatic
Duschl., and A. C. Cheng.: ‘A wearable smartphone-based classification of heartbeats using ecg morphology and
platform for real-time cardiovascular disease detection via heartbeat interval features’, IEEE Trans Biomed Eng., 2004,
electrocardiogram processing’, IEEE trans. Inf.Technol. 51, (7), pp. 1196-1206
Biomed., 2010, 14, (3), pp. 734-740 [23] N. Dogru and A. Subasi.: ‘Traffic accident detection
[4] J. A. Nasiri., M. Naghibzadeh., H. S. Yazdi., and B. using random forest classifier’, IEEE 15th Learning and
Naghibzadeh.: ‘Ecg arrhythmia classification with support Technology Conference (L&T)., 2018, pp. 40-45
vector machines and genetic algorithm’, IEEE third UK [24] Ari S., Hembram K.,Saha g.: ‘Detection of cardiac
Sim European Symposium on Computer Modelling and abnormality from PCG signal using LMS based least square
Simulation., 2009, pp. 187-192 SVM classifier’, Expert Syst. Appl., 2010, 37, (12), pp.
[5] D. A. Coast., R. M. Stern., G. G. Cano., and S. A. Briller.: 8019-8026
‘An approach to cardiac arrhythmia analysis using hidden [25] Stockwell R., Mansinha L., Lowe R.: ‘Localization of the
markov models’, IEEE Trans Biomed Eng., 1990,37, (9), pp. complex spectrum: the S transform’, IEEE Trans. Signal
826-836 Process., 1996, 44, (4), pp. 998-1001
[6] Y. H. Hu., S. Palreddy., W. J. Tompkins.: ‘A patient-
adaptable ecg beat classifier using a mixture of experts
approach’, IEEE Trans Biomed Eng., 1997, 44, (9), pp. 891-
900
[7] M. K. Das., S. Ari.: ‘Patient-specific ecg beat classification
technique’, Healthcare technology letters., 2014, 1, (3), pp.
98-103
[8] L. Breiman.: ‘Random Forests’, Machine Learning., 2001,
45, (1), pp. 05-32
[9] G. B. Moody., R. G. Mark.: ‘The impact of the mit-bih
arrhythmia database,’ IEEE Engineering in Medicine and
Biology Magazine., 2001, 20, (3), pp. 45-50
[10] J. Pan and W., J. Tompkins.: ‘A Real-Time QRS Detection
Algorithm’, IEEE Trans Biomed Eng., 1985, 32, (3), pp.
230-236
[11] M. Thomas., M. K. Das.,S. Ari.: ‘Automatic ecg arrhythmia
classification using dual tree complex wavelet based
features’, AEU International Journal of Electronics and
Communications., 2015, 69, (4), pp. 715-721
[12] Haykin S.: ‘Neural Networks’(Pearson Education Asia,
New Delhi,2009)
[13] S. Bernard., L. Heutte and S. Adam.: ‘On the selection
of decision trees in Random Forests’, International Joint
Conference on Neural Networks (IJCNN), Atlanta, GA.,
2009, 1109,(10), pp. 302-307

Healthcare Technology Letters, pp. 5–5 5


The Institution of Engineering and Technology 2018
c

Вам также может понравиться