Available online at www.sciencedirect.com
ScienceDirect
Available online at www.sciencedirect.com
Available online Available onlineat at www.sciencedirect.com www.sciencedirect.com Procedia Computer Science 00 (2018) 000–000 Available online at www.sciencedirect.com
ScienceDirect
ScienceDirect ScienceDirect Procedia Computer Science 00 (2018) 000–000 ScienceDirect
www.elsevier.com/locate/procedia
Procedia Computer Science Procedia Science00 132(2018) (2018)000–000 606–613and Datawww.elsevier.com/locate/procedia International Conference onComputer Computational Intelligence Science (ICCIDS 2018) Procedia Computer Science 00 (2018) 000–000 www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia
Ischemia and Conference Arrhythmia Classification Using Time-Frequency Domain International on Computational Intelligence and Data Science (ICCIDS 2018) International Conference on Computational Intelligence and Data Science (ICCIDS 2018) Features of QRS Complex International Conference on Intelligence and Data Science (ICCIDS 2018) Ischemia and ArrhythmiaComputational Classification Using Time-Frequency Domain Ischemia and Arrhythmia Classification Using Time-Frequency Domain b of Sonam QRS Complex Ischemia and Arrhythmia Classification Using Time-Frequency Domain Akash Kumar Features Bhoia* , Karma Sherpa , Bidita Khandelwalc Features of QRS Complex Features of QRS Complex a* b c Department of Electrical & Electronics Engineering, Sikkim Manipal Institute of Technology (SMIT), Sikkim Manipal University, India a* ,b
Akash Kumar Bhoi , Karma Sonam Sherpa , Bidita Khandelwal Department General Medicine,Central Referral Hospital and SMIMS,bSikkim Manipal University, Indiac Akash Kumar Bhoia* a*, Karma Sonam Sherpa b, Bidita Khandelwal c Kumar Bhoi , Karma Sonam Sherpa , Bidita Department ofAkash Electrical & Electronics Engineering, Sikkim Manipal Institute of Technology (SMIT),Khandelwal Sikkim Manipal University, India c
a* ,b
c
Department General Medicine,Central Referral Hospital and SMIMS, Sikkim (SMIT), ManipalSikkim University, India Abstract a* ,b Department of Electrical & Electronics Engineering, Sikkim Manipal Institute of Technology Manipal University, India a* ,b
c
Department General Medicine,Central Referral Hospital and SMIMS, Sikkim Manipal University, IndiaUniversity, India Department of Electrical & Electronics Engineering, Sikkim Manipal Institute of Technology (SMIT), Sikkim Manipal
c Department wave General Referral Hospital and SMIMS, Sikkim Manipal University, India If the synchronization The QRS complex is the significant ofMedicine,Central electrocardiogram (ECG) and occurs during ventricular depolarization. Abstract(i.e., depolarization) occurs between endocardial cardiomyocytes and outer layers, then abnormal morphological changes in QRS problem complex take place which is possible in case of arrhythmia and ischemia. The proposed approach describes time domain measures of such Abstract The QRS complex the of significant wave (ECG) of andQRS occurs duringalong ventricular If thefeat synchronization Abstract alterations (i.e., the is ratio average rise & of fallelectrocardiogram amplitude and interval) complex with thedepolarization. frequency domain ure i.e., peak problem occurs between endocardial cardiomyocytes outerduring layers, then abnormal morphological in QRS frequency anddepolarization) power of significant mean QRSwave complexes. An improvised Difference Operation Method (DOM) (Yeh Yun-Chi et al., 2008) is The QRS(i.e., complex is the of electrocardiogram (ECG) andand occurs ventricular depolarization. If thechanges synchronization complex take place is possible in case arrhythmia and (e.g., ischemia. The proposed describes time of QRS such implemented with added likewave preprocessing techniques baseline drifts and approach noise The domain proposed methodology is The QRS(i.e., complex iswhich the features significant of of electrocardiogram (ECG) andand occurs during ventricular depolarization. If themeasures synchronization problem depolarization) occurs between endocardial cardiomyocytes outer layers, thencancellation). abnormal morphological changes in alterations (i.e., the standard ratio & fall amplitude and interval) of QRS complex along with the frequency domain feat ure i.e., peak evaluatedtake with the databases, FANTASIA (healthy subjects), MIT-BIH Arrhythmia database (MITDB), andmeasures European problem (i.e., depolarization) occurs rise between endocardial cardiomyocytes and outer layers, then abnormal morphological changes in QRS complex place whichofisaverage possible ini.e., case of arrhythmia and ischemia. The proposed approach describes time domain of ST-T such frequency andplace power ofofmean QRS An improvised Difference Method (DOM) (Yeh etarrhythmic al., i.e., 2008) is database (EDB) respectively. Linear Discriminant Analysis (LDA) and decision tree are carried out for healthy, and complex take which isaverage possible incomplexes. case arrhythmia and ischemia. The Operation proposed approach describes time Yun-Chi domain measures of peak such alterations (i.e., the ratio rise & fallofamplitude and interval) of QRS complex along with theclassification frequency domain feat ure implemented with added features like preprocessing techniques (e.g., baseline drifts and noise cancellation). The proposed is ischemic subjects the time-frequency domain characteristics. The Operation Naive Bayes’ Classifiers also implemented where two alterations (i.e., theincorporating ratio average rise & fall amplitude and interval) of QRS complex along with the frequency domain feat ure these i.e., peak frequency and power of of mean QRS complexes. An improvised Difference Method (DOM) (Yeh Yun-Chi etmethodology al., 2008) is evaluated with the standard databases, i.e., FANTASIA (healthy subjects), MIT-BIH Arrhythmia database (MITDB), and et European ST-T classes, i.e., ischemia and arrhythmia visually distinguished by(e.g., convergently spreading in different areas. Real-time hardware validation of frequency and power of features mean QRS complexes. An improvised Difference Operation Method (DOM) (Yeh al., 2008) is implemented with added likeare preprocessing techniques baseline drifts and noise cancellation). The Yun-Chi proposed methodology is database (EDB) respectively. Linear Discriminant (LDA)(e.g., and decision tree are carried out for classification healthy, arrhythmic and this approach has also executed. The future scope Analysis of techniques this(healthy approach is to be validated with ischemia-induced arrhythmia subjects for precise implemented added features like preprocessing baseline drifts and noise cancellation). The proposed methodology is evaluated withwith the standard databases, i.e., FANTASIA subjects), MIT-BIH Arrhythmia database (MITDB), and European ST-T ischemic subjects incorporating the time-frequency domain characteristics. The Naive also(MITDB), implemented theseST-T two identification classification. evaluated withand the standard databases, i.e., FANTASIA (healthy subjects), MIT-BIH Arrhythmia database and where European database (EDB) respectively. Linear Discriminant Analysis (LDA) and decision tree areBayes’ carriedClassifiers out for classification healthy, arrhythmic and classes, ischemia and arrhythmia are visually distinguished by convergently in different Real-time hardware validation of databasei.e., (EDB) respectively. Linear Discriminant Analysis and decision tree areBayes’ carried out forareas. classification healthy, arrhythmic and ischemic subjects incorporating the time-frequency domain (LDA) characteristics. The spreading Naive Classifiers also implemented where these two this approach has incorporating alsoPublished executed. The future scopedistinguished ofdomain this approach is to beThe validated arrhythmia subjects for precise ischemic subjects the time-frequency characteristics. Naive with Bayes’ Classifiers implemented where these two © 2018 The Authors. by Elsevier B.V. classes, i.e., ischemia and arrhythmia are visually by convergently spreading inischemia-induced different areas.also Real-time hardware validation of identification and classes, i.e., ischemia and arrhythmia are visually spreading different areas. Real-time hardware validation of Peer-review under responsibility of scientific committee of by theconvergently International Conference on Computational Intelligence and Data this approach has classification. also executed. Thethe future scopedistinguished of this approach is to be validated withinischemia-induced arrhythmia subjects for Science precise this approach has classification. also executed. The future scope of this approach is to be validated with ischemia-induced arrhythmia subjects for precise (ICCIDS 2018). identification and © 2018 The Authors. Published by Elsevier B.V. identification and classification. © 2018The The Authors. Published by Ltd. Keywords: QRS complex, DOM, LDA, Tree, Bayes’ Classifiers, FANTASIA, Conference MITDB, EDB.on Computational Intelligence and Data Science Peer-review under responsibility theElsevier scientific committee of the International © 2018 Authors. Published byofDecision Elsevier B.V.Naive (ICCIDS 2018). This is an open article by under CC license (https://creativecommons.org/licenses/by-nc-nd/3.0/) © 2018 The Authors. Published B.V.BY-NC-ND Peer-review underaccess responsibility ofElsevier the the scientific committee of the International Conference on Computational Intelligence and Data Science Peer-review underresponsibility responsibility of scientific the scientific committee the International Conference on Computational and Peer-review under of the committee of the of International Conference on Computational Intelligence Intelligence and Data Science (ICCIDS 2018). Keywords: QRS complex, DOM, LDA, Decision Tree, Naive Bayes’ Classifiers, FANTASIA, MITDB, EDB. (ICCIDS 2018). 1. Introduction Data Science (ICCIDS 2018).
Keywords: QRS complex, DOM, LDA, Decision Tree, Naive Bayes’ Classifiers, FANTASIA, MITDB, EDB. Keywords: QRS complex, DOM, Decision component Tree, Naive Bayes’ Classifiers, MITDB, EDB. The QRS complex is LDA, a pivotal of wave in FANTASIA, any normal ECG beat.
Depolarization in QRS morphology, 1. Introductionto its duration and amplitude; denotes many cardiac malfunctions [1]. D Romero et al., [2, 3] had well evaluated corresponding 1. theIntroduction changes in deploraziation of QRS complex during acute myocardial ischemia. They had further established that information 1. The is a inpivotal component of wave in any normal of ECG Depolarization in QRS morphology, onIntroduction QRS QRS slope complex could be used monitoring myocardial ischemia [4]. Slowing interbeat. – myocardial conduction leading to changes corresponding to its duration and amplitude; denotes many cardiac malfunctions [1]. D Romero et al., [2, 3] had well evaluated The QRS complex is a pivotal component of wave in any normal ECG beat. Depolarization in QRS morphology, in QRS morphology in some animals during ischemia have been demonstrated [5]. Therefore, changes in the depolarization The QRSincomplex is a and pivotal component of many wave in any malfunctions normal ECG[1]. beat. Depolarization in QRS morphology, the changes deploraziation of QRS complex during acute myocardial ischemia. They had further established that information corresponding to its duration amplitude; denotes cardiac D Romero et al., [2, 3] had well evaluated phase of the QRS complex may lead to additional information for analysis beyond ST-T segment. Miyake C.Y. et al., suggested corresponding to its be duration amplitude; denotes many cardiac malfunctions [1]. Romero etconduction al., [2, 3] had well on could used deaths inand myocardial ischemia [4]. inter –Dmyocardial toevaluated changes the changes deploraziation ofmonitoring QRS during acute myocardial ischemia. They had established information thatQRS thereslope areinsudden infant duecomplex to cardiac failure before theSlowing onset ofofarrhythmia andfurther the markers are leading thethat changes in ST the changes incould deploraziation ofanimals QRS complex during ischemia acute myocardial ischemia. They had further established that information in QRS morphology in some during ischemia have been demonstrated [5]. Therefore, changes in the depolarization on slope be used in monitoring myocardial [4]. Slowing of inter – myocardial conduction leading to changes and widening of QRS complex [6]. Moreover, evaluation of QRS angles (ØR) will lead to the severity of acute myocardial on QRS QRS could beinused in animals monitoring myocardial ischemia [4].analysis Slowingbeyond of inter – myocardial conduction leading tosuggested changes phase of slope the QRS complex may lead to during additional information for ST-T segment. Miyake C.Y. al., in morphology some ischemia have been demonstrated [5]. Therefore, in et theet depolarization ischemia. A real-time machine implantable algorithm was proposed by Pan and Tompkins [7] andchanges C. Saritha al., developed a in QRS morphology in some animals during ischemia have been demonstrated [5]. Therefore, changes in theet changes depolarization that there are sudden infant deaths due to cardiac failure before the onset of arrhythmia and the markers are in ST phase of the QRS complex may lead to additional information for analysis beyond ST-T segment. Miyake C.Y. al., suggested suitable MATLAB simulator using wavelet transform [8] for analysis of QRS morphology. phase of the complex may lead to additional information analysis beyond ST-T segment. Miyake C.Y. etchanges al.,myocardial suggested and widening of QRS complex [6].due Moreover, of for QRS anglesof (ØR) will lead themarkers severityare ofthe acute that there areQRS sudden infant deaths to cardiacevaluation failure before the onset arrhythmia andtothe in ST that widening there A arereal-time sudden deaths to cardiac failure the by onset arrhythmia and are the changes in STa ischemia. machine implantable algorithm wasbefore proposed Panof(ØR) and Tompkins [7] and C. Saritha al., developed and of QRSinfant complex [6].due Moreover, evaluation of QRS angles will lead tothe themarkers severity ofet acute myocardial and widening of QRS complex [6].wavelet Moreover, evaluation of QRS by angles (ØR) will lead[7] to and the C. severity myocardial suitable MATLAB simulator using transform [8] analysis ofPan QRSand morphology. ischemia. A real-time machine implantable algorithm wasfor proposed Tompkins Sarithaofetacute al., developed a ischemia.MATLAB A real-time machine implantable algorithm [8] wasforproposed Tompkins [7] and C. Saritha et al., developed a suitable simulator using wavelet transform analysisby of Pan QRSand morphology. suitable MATLAB simulator using wavelet transform [8] for analysis of QRS morphology. *Akash Kumar Bhoi, Tel.: +91-8001822676 E-mail address:
[email protected];
[email protected]
1877-0509 ©2018 2018The TheAuthors. Authors. Published by Elsevier *Akash Kumar Bhoi, Tel.: +91-8001822676 1877-0509© Published by Elsevier B.V.Ltd. E-mail
[email protected];
[email protected] This isaddress: an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/3.0/) Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data Science *Akash Kumar Bhoi, Tel.: +91-8001822676 Peer-review under responsibility of
[email protected] scientific committee of the International Conference on Computational Intelligence and Data Science (ICCIDS 2018). E-mail address:
[email protected]; *Akash Kumar Bhoi, Tel.: +91-8001822676 (ICCIDS 2018). 1877-0509© 2018 The Authors. Published by Elsevier B.V. E-mail address:
[email protected];
[email protected] 10.1016/j.procs.2018.05.014 Peer-review under responsibility of the scientific committee 1877-0509© 2018 The Authors. Published by Elsevier B.V. of the International Conference on Computational Intelligence and Data Science (ICCIDS 2018). 1877-0509© 2018 The Authors. Published by Elsevier B.V. of the International Conference on Computational Intelligence and Data Science Peer-review under responsibility of the scientific committee Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data Science (ICCIDS 2018). (ICCIDS 2018).
Akash Kumar Bhoi et al. / Procedia Computer Science 132 (2018) 606–613
Akash Kumar Bhoi / Procedia Computer Science 00 (2018) 000–000
607
Various sources have also implemented wavelet transform for QRS detection [9–14]. Other complex methods like neural network [10,15], K-Nearest Neighbor [16] Genetic Algorithms [17], Hidden Markov Models [18], Support Vector Machines [19], and combination of different algorithms for morphological analysis of QRS complex was also described [20, 21, 22]. The primary challenge lies in analyzing QRS complex to extract important features (i.e., rise amplitude, fall amplitude and interval) and classify ischemia, arrhythmia and normal subjects using decision tree and Linear Discriminant Analysis (LDA) classifier. Moreover, the frequency domain features which already been validated were incorporated along with time domain features to improvise the classification performance. Consequently, only selected sets of 108 recordings are viewed for evaluation of this methodology. 2. Methodology 2.1. ECG Datasets Some standard ECG datasets have been approached for the extraction of time domain features, classifications and successful detection of the QRS complex. The selected data from the MIT-BIH Arrhythmia Database (MITDB) [23], FANTASIA [24] and European ST-T database (EDB) [23] respectively. The sampling rate of MITDB, FANTASIA and EDB are 360Hz, 250Hz and 250 Hz respectively. This method has been proven over 1 hour (selected output length of recording using PhysioBank ATM [23, 24]) duration of recording from each database. To perform uniform classification analysis using LDA and decision tree, 108 selected episodes having 100% accuracy in QRS complex detection are evaluated. The entire number of beats (QRS complex) obtained from these three databases is 10100 and these beats are validated with the proposed methodology for QRS complex detection as well as features extractions. 2.2. QRS Complex Detection As here FANTASIA, EDB and MITDB databases are considered which carries high frequency and abrupt morphological features. The improvisation of earlier DOM for QRS complex detection has been implemented where preprocessing steps has been added and both high-low frequency components also eliminated from the input ECG signals. In the preprocessing stage, the input ECG signal is passed through the noise and baseline drift cancellation processes. Moving Average Filtering (MAF) [25] is implemented for removal of baseline drift and is given by the following difference equation (1);
where, and
= smoothed value for the = span.
sample point,
= number of neighboring data points on both side of
This approach is followed the rules of Curve Fitting Toolbox such as: (i) odd span (ii) Smoothed data point at the span center (iii) specified number of neighbors on either side cannot be accommodated for the adjusted data points of the span (iv) smoothed end points. And so the high frequencies (above 100 Hz) components are eliminated by using a low-pass filtering where as the FFT and IFT are performed to get rid of low frequency subcomponents (below 0.05 Hz) [26] and restored the ECG signal. Fourier transform for sampled function in discrete form using equation (2):
Inverse Fourier transform with discrete counterpart using equation (3):
Based on medical definition [27] and the author Yeh Yun-Chi et al., [28], threshold is defined for R-peak and the search interval is performed to detect the correct R-peaks which are either the extreme positive or negative values within the interval [28]. “Difference Operation Method (DOM)” was implemented for detection of Q and S points as well. The elaborate description of this improvised DOM was presented in the author’s earlier work [29]. The detected QRS complex is shown in Fig. 1, 2.
Akash Kumar Bhoi et al. / Procedia Computer Science 132 (2018) 606–613
608
Akash Kumar Bhoi / Procedia Computer Science 00 (2018) 000–000
Detection Performance (P) [30] calculated using equation (4), of the proposed method for the selected signals of the three databases;
where, ΣS = total number detected beats (i.e., 10100), ΣM = detected missing beats of QRS complex and ΣF = detected false beat of the QRS complex.
Fig. 1. Filtered ECG Signal showing QRS complex (FANTASIA data # f1o01)
Fig. 2. Filtered ECG Signal showing QRS complex (MITDB data # 106)
# 10 sec data are plotted in Fig.1 and Fig.2 for better visualization 3. QRS Features Extraction 3.1. Ratio of Average Rise & Fall Amplitude The Average Rise Amplitude ( computed using (5) and (6);
) and Average Fall Amplitude (
) of QRS complex can be
[where, abs= absolute value, R= R-peaks, Q= Q points, S= S points of ECG signals] is calculated by subtracting the mean absolute value of R-peaks and Q points whereas, the is formulated by subtracting the mean absolute value of S points and R-peaks points. The ratio of Average Rise and Fall Amplitude is calculated using (7); The
Akash Kumar Bhoi / Procedia Computer Science 00 (2018) 000–000
Akash Kumar Bhoi et al. / Procedia Computer Science 132 (2018) 606–613
609
3.2. QRS complex Interval The FANTASIA, MITDB and EDB signals are having length of 1 hour. After successful detection of QRS complexes, the signal lengths of the mean QRS complexes are multiplied with the sampling intervals (i.e., 0.004 Sec for FANTASIA & EDB and 0.0027 Sec for MITDB) to obtain the QRS interval. Table.1 Time domain features of QRS complex
Ratio Interval Ratio Fantasia MITDB R&F (Sec) R&F f1o01 -1.0600 0.1160 100 -1.0737 f1o02 -0.9244 0.0880 101 -0.9658 f1o04 -0.9340 0.0800 105 -0.9400 f1o05 -0.8324 0.0600 106 -0.9228 f1o06 -0.9856 0.1440 109 -0.7569 f1o07 -0.9968 0.1040 111 -0.8201 f1o09 -0.9727 0.1720 112 -0.8084 f1o10 -0.9830 0.0800 113 -0.9338 f1y01 -0.9955 0.0720 114 -0.7778 f1y02 -0.9266 0.0760 115 -0.7809 f1y03 -0.9337 0.0640 116 -0.9047 f1y04 -0.7506 0.1200 119 -1.0022 f1y05 -1.0012 0.1200 121 -0.9973 f1y06 -0.9311 0.0600 122 -1.0556 f1y07 -1.0043 0.0960 123 -1.0556 f1y08 -1.0312 0.0640 124 -0.9354 f1y09 -0.9800 0.1480 200 -0.8978 f1y10 -0.9128 0.0640 202 -0.9214 f2o01 -0.8751 0.0680 203 -0.8488 f2o02 -0.8934 0.0712 205 -1.0439 f2o03 -1.1323 0.1080 208 -0.9133 f2o04 -0.8965 0.0720 210 -0.9003 f2o06 -0.9311 0.0600 212 -0.7195 f2o07 -0.9752 0.1040 213 -0.8247 f2o09 -0.9238 0.0760 214 -0.9247 f2o10 -0.9830 0.0800 215 -0.6107 f2y01 -0.7150 0.0640 220 -0.7543 f2y02 -0.9402 0.0960 221 -0.9283 f2y03 -0.8466 0.0640 222 -0.9868 f2y04 -0.8717 0.0960 223 -0.9610 f2y05 -0.9937 0.1680 228 -1.1905 f2y06 -0.9113 0.1240 230 -0.6256 f2y07 -1.0928 0.0880 231 -0.7901 f2y08 -0.9786 0.1040 232 -0.6598 f2y09 -1.0431 0.1080 233 -0.7082 f2y10 -0.9896 0.0480 234 -0.9762 # Ratio R&F= Ratio of Average Rise & Fall Amplitude Interval (sec)= QRS complex Interval in second
Interval
EDB
0.0340 0.0760 0.1280 0.0380 0.1260 0.1100 0.0760 0.0380 0.0200 0.0600 0.0360 0.1100 0.0980 0.0780 0.0780 0.1080 0.0880 0.1200 0.1380 0.0820 0.0920 0.1220 0.0860 0.0960 0.1220 0.0480 0.0320 0.1460 0.0620 0.0960 0.0740 0.0940 0.0540 0.0940 0.0640 0.0860
e0103 e0104 e0105 e0108 e0115 e0116 e0118 e0119 e0121 e0122 e0123 e0124 e0125 e0126 e0127 e0133 e0136 e0139 e0147 e0148 e0151 e0154 e0155 e0159 e0161 e0162 e0163 e0166 e0170 e0202 e0203 e0204 e0206 e0207 e0208 e0210
Ratio R&F -0.9944 -1.5632 -0.7512 -0.8022 -0.9549 -0.8946 -0.8874 -0.8992 -0.9060 -0.9491 -0.7322 -0.7168 -0.7282 -0.7373 -0.8756 -0.8380 -0.7041 -0.6151 -0.9934 -0.9993 -0.8469 -0.7995 -0.8294 -0.8847 -0.7224 -0.5544 -0.5439 -1.1327 -0.7993 -0.6536 -0.9004 -0.9519 -0.9083 -0.9800 -0.6946 -0.9621
Interval (Sec) 0.1400 0.1880 0.0880 0.0840 0.1400 0.1720 0.0760 0.0560 0.0640 0.1160 0.0760 0.0880 0.0960 0.1000 0.0720 0.0960 0.0400 0.1000 0.0960 0.0960 0.0720 0.0720 0.1000 0.1760 0.0800 0.0720 0.0800 0.0800 0.0640 0.1040 0.1640 0.1560 0.0760 0.1560 0.0880 0.1720
4. Classification using Linear Discriminant Analysis (LDA) and Decision Tree Here and the Interval of QRS complex of subjects from all three databases are considered. The three databases are supposed to be classified based on these two features (i.e., and Interval) of QRS complex (Healthy ‘HLY’, Arrhythmic ‘ARY’, Ischemic ‘ISH’). The misclassification error (i.e., resubstitution error) (Fig. 3), of LDA [31] on observations with known class labels (training set) is calculated as 12.7%. Whereas, resubstitution error in case of Quadratic Discriminant Analysis (QDA) is computed as 9.17%, this is comparatively less than LDA misclassification. Here, the decision tree is being used for classifying these three classes with two features (time domain) and setting the best tree to establish the difference between two or more classes by operation search and decision analysis approach. The re-substitution error is 4.15% of decision trees. These misclassification results of LDA and QDA are marginally higher as compared to decision tree, which in this case found to be promising for classification. The performance of LDA is validated with the time-frequency domain features
610
Akash Kumar Bhoi et al. / Procedia Computer Science 132 (2018) 606–613
Akash Kumar Bhoi / Procedia Computer Science 00 (2018) 000–000
(i.e., , Interval and Peak Frequency, Power) and the misclassification error are calculated as 7.67% whereas in case of QDA is found to be off 5.78%. The features are not so convincing factors for classifying arrhythmia and ischemia, because of the mean and variance. Still the probability of classification supposed to be validated taking time domain features of QRS (i.e., the ratio of average rise & fall amplitude and interval) and from Fig.4, the major classification can also be seen in the case of healthy and ischemic subjects.
Fig. 3. LDA misclassification
Fig. 4. Classification using LDA for three classes
Fig. 5. Decision Tree Classification with four features
As the error rate is still high in case of LDA and QDA, the decision tree classifier is found to be suitable (as already seen better results using time domain features) and the re-substitution error is 2.15% for these three classes (Fig. 5).
Akash Kumar Bhoi / Procedia Computer Science 00 (2018) 000–000
Akash Kumar Bhoi et al. / Procedia Computer Science 132 (2018) 606–613
611
5. Naive Bayes’ Classifiers Naive Bayes’ Probability Classifiers [32] is oftentimes employed for examining the probability of information categorization. This is implemented for visualizing the classification probability and evaluating the results obtained using LDA and decision classifier. The Naive Bayes’classified result (Fig. 6), also predicting the same possibility that was obtained with the applied approaches. As per the literature study, there are complex cardiovascular cases of ischemic induced arrhythmia conditions with transition stage, then it is very important to distinguish between ischemic and arrhythmia subjects. Moreover, with this plot (Fig. 6), a probability also rises to investigate the overlapping part of both the classes and evaluate possible the conversion stage.
Fig. 6. Naive Bayes’ Probability Classifiers (Classification Probability)
6. Discussion and Performances Comparison Takeshi Tsutsumi et al., has suggested that there has been unclearity in the findings associated with Distribution of Frequency Power (DFP) within the QRS complex. The Distribution of Frequency Power (DFP) within the QRS complex in Ischemic Cardiomyopathy (ICM) with Lethal Ventricular Arrhythmias (L-VA) was investigated using Continuous Wavelet Transform (CWT). It has been found that, 180-190 Hz abnormal frequency power provides arrhythmogenic signals in ICM with LVA that may be affiliated with the fibrous tissue proliferation [33]. Moreover, Takeshi Tsutsumi et al. [34], also indicated that time– frequency analysis of the QRS complex has not been uniformly accepted. It has been found that, the Integrated Time–Frequency Power (ITFP) values were lower in Mid-Frequency QRS (MF-QRS, 15–80 Hz) patients with anterior or inferior Myocardial Infarction (MI). However, ITFP values were significantly greater in Low-Frequency QRS (LF-QRS, 5–15 Hz) and HighFrequency (HF-QRS, 150–250 Hz) bands in Ischemic Cardiomyopathy (ICM) patients than in other groups. Feng Jin et al., has characterized the fragmented QRS (fQRS) complex, which indicates the abnormal electrical activation. Patients with heart disease and subjected to Sudden Cardiac Death (SCD) are with possible potential risk with fQRS. The fQRS to be quantified using Intrinsic Time-scale Decomposition (ITD) method considering precordial ECG leads, V1–V6 [35]. Here, the classification performance could have been better with bigger datasets. The ratio of average rise and fall amplitude of the QRS complex is computed and tabulated in Table.1. The mean for FANTASIA is -0.948581 mV, MITDB is -0.886581 mV and EDB is -0.852953 whereas the mean Interval of QRS complex for healthy subject is 0.0924222 sec, arrhythmic subject is 0.0836111 sec and ischemic subject is 0.102667 sec. Using statistical classification techniques, i.e., LDA and decision tree the same features are taken and the three different classes, e.g. healthy, arrhythmic and ischemic are classified. The misclassification error is quite less in cases of decision tree as compared to the LDA and QDA. The complete decision tree is developed with decision rule and class assignments using time domain characteristics. The resubstitution error for LDA and QDA found to be 8.19% and in the case of decision tree it was off 2.67% for healthy and ischemic subjects. Looking classification result of LDA and decision tree, Naive Bayes’ Probability Classifiers was implemented and the prominent classification (Fig. 5), was achieved for ischemic and healthy subjects. The classification performance of ischemia and arrhythmia using frequency domain features was already examined in previous work [29] and the misclassification error using decision tree in that case found to be 8.33%. For improvising this proposed method, the frequency domain features, i.e., peak frequency (Peak Freq) and power (Pwr) of mean QRS complexes [29] are merged along with the Ratio R&F and Interval of QRS complexes. The decision tree classification (Fig. 5), is performed by considering these four features (time-frequency domain features) and the misclassification rate is found to be 1.58 % for healthy and arrhythmia and 1.26 % for healthy and ischemia subjects. The overall classification error rate (between healthy, ischemia and
612
Akash Kumar Bhoi et al. / Procedia Computer Science 132 (2018) 606–613
Akash Kumar Bhoi / Procedia Computer Science 00 (2018) 000–000
arrhythmia subjects) also significantly reduced to 2.15%. The better classification is observed in case of ischemia and arrhythmia groups where, the misclassification error found to be 1.89%. The complete algorithm was developed in the MATLAB platform. The overall analysis has been carried out to assess the performance among LDA, QDA and decision tree by taking time, time-frequency domain features and the distinct results are held in each case with implying a scope for improvisation. Real-time algorithm (i.e., detection, feature extraction and classification) validation has been performed in Biomedical Instrumentation Laboratory of SMIT. The ECG signal is acquired using an ECG machine (model: ST2351) and processed by computer via NI-Elvis-II DAQ, where Semi Real-Time analysis has been done on MatlabR2016a platform. Real-time PC-based ECG system can be upgraded with add-on tools with MATLAB for cardiovascular abnormality detection with other possible features into account. 7. Conclusions The proposed methodology not only classifies the different cardiovascular conditions (i.e., ischemia and arrhythmia) but also accurately detect the inflection points of QRS complexes. Improvement in classification can be distinguishable observed after considering frequency domain features along with time domain characteristics. The QRS complex amplitude, duration and peak frequency and power change can be marked for arrhythmic and ischemic condition detection. There will be certain changes in rise & fall amplitude and interval of tachyarrhythmic QRS complex. The literature survey shows that ST-T segment analysis was ideal for ischemia detection and analysis, but the proposed methodology takes the new challenge of detecting and analyzing the depolarizing wave (QRS complex) of the electrocardiographic signals. Here classification results of the decision tree and Naive Bayes’ Classifier also shows prominent two different classes’ i.e., ischemic and arrhythmia subjects. The further search involves with the analytic thinking of time-frequency domain QRS complex changes during ventricular tachyarrhythmic events and its classification. References [1] Z.E.H. Slimanea, A.N. Alib, (2009) “QRS complex detection using Empirical Mode Decomposition”, Digital Signal Processing, Elsevier Inc20(4): 1221– 1228. [2] D Romero, M Ringborn, P Laguna, O Pahlm, E Pueyo, (2010) “A Vectorial Approach for Evaluation of Depolarization Changes during Acute Myocardial Ischemia”, Computing in Cardiology37: 265−268. [3] D. Romero, M. Ringborn, P. Laguna, E. Pueyo, (2013) “Detection and quantification of acute myocardial ischemia by morphologic evaluation of QRS changes by an angle-based method”, Journal of Electrocardiology, Elsevier Inc46: 204–214. [4] E Pueyo, L Sornmo, P Laguna, (2008) “QRS Slopes for Detection and Characterization of Myocardial Ischemia”, IEEE Transactions on Biomedical Engineering 55(2). [5] R. P Holland, H. Brooks, (1976) “The QRS complex during myocardial ischemia: An experimental analysis in the porcine heart”, J. Clin. Invest.57: 541– 550. [6] C.Y. Miyake, A.M. Davis, K.S. Motonaga, A.M. Dubin, C.I. Berul, F. Cecchin, (2013) “Infant Ventricular Fibrillation After ST-Segment Changes and QRS Widening A New Cause of Sudden Infant Death?”, Circ Arrhythm Electrophysiol 6(4): 712-8. [7] J. Pan, W.J. Tompkins, (1985) “A real-time QRS detection algorithm”, IEEE Trans. Biomed. Eng. BME32 (3): 230–236. [8] C. Saritha, V. Sukanya, Y. N. (2008) “Murthy, ECG Signal Analysis Using Wavelet Transforms”, Bulg. J. Phys. 35: 68–77. [9] C. Li, C. Zheng, C. Tai, (1995) “Detection of ECG characteristic points using wavelet transforms”, IEEE Trans. Biomed. Eng.42 (4): 21–28. [10] B. Abibullaev, H. D. Seo, (2011) “A new QRS detection method using wavelets and artificial neural-networks”, Springer J.Med. Syst.35: 683–691. [11] G. Ruchita, A.K. Sharma, (2010) “Detection of QRS complexs of ECG recording based on wavelet transform using Matlab”, Int. J.Eng. Sci.2 (7): 3038– 3034. [12] A.N. Dinh, D.K. Kumar, N.D. Pah, P. Burton, (2002) “Wavelet for QRS detection, in: Engineering in medicine and Biology society”, 23rd Conference of IEEE, 7803–7211. [13] Z. Zidelmal, A. Amirou, M. Adnane, A. Belouchrani, (2012) “QRS detection using wavelet coefficients”, Comput. Method. Program. Biomed. 107 (3): 490– 496. [14] S.W. Chen, C.H. Chen, H.L. Chan, (2006) “A real-time QRS method based on moving-averaging incorporating with wavelet denoising”, Comput. Method. Program. Biomed. 82 (3): 187–195. [15] Q. Xue, Y.H. Hu, J. Tompkins, (1992) “Neural-network-based adaptive matched filtering for QRS detection”, IEEE Trans. Biomed. Eng. 39 (4): 317–329. [16] I. Saini, D. Singh, A. Khosla, (2013) “QRS detection using K-Nearest Neighbour algorithm (KNN) and evaluation on standard ECG databases”, J. Adv. Res. 4(4): 331–344. [17] R. Poli, S. Cagnoni, G. Valli, (1995) “Genetic design of optimum linearand nonlinear QRS detectors”, IEEE Trans. Biomed. Eng. 42(11): 1137–1141. [18] D.A. Coast, R.M. Stern, G.G. Cano, S.A. Briller, (1990) “An approach tocardiac arrhythmia analysis using hidden Markov models”, IEEE Trans. Biomed. Eng. 37 (9): 826–836. [19] S.S. Mehta, N.S. Ligayat, (2007) “Comparative study of QRS detectionin single lead and 12-lead ECG based on entropy and combined entropy criteria using support vector machine”, J.Theor. Appl. Inform. Technol.3 (2): 8–18. [20] C. Meyer, J.F. Gavela, M. Harris, (2006) “Combining algorithms in automatic detection of QRS complexes in ECG signals”, IEEE Trans. Inform. Technol. 10 (3): 468–475. [21] J.C.T.B. Moraes, M. Seixas, F.N. Vilani, E.V. Costa, (2002) “A QRS complex detection algorithm using electrocardiogram leads”, Comput. Cardiol.29: 205–208. [22] Z. Zidelmal, A. Amirou, D. Ould-Abdeslam, A. Moukadem, and A. Dieterlen. (2014) “QRS detection using S-Transform and Shannon Energy”, Computer Methods and Programs in Biomedicine116(1): 1–9. [23] A.L. Goldberger, L.A.N. Amaral, L. Glass, J.M. Hausdorff, P.C. Ivanov, R.G. Mark, J.E. Mietus, G.B. Moody, C.K. Peng, H.E. Stanley, (2000) “PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals”, Circulation 101(23): e215-e220. [24] N. Iyengar, C.K. Peng, R. Morin, A.L. Goldberger, L.A. Lipsitz, (1996) “Age-related alterations in the fractal scaling of cardiac interbeat interval dynamics”, Am J Physiol271: 1078-1084. [25] M. Kaur, B. Singh, Seema, (2011). “Comparison of Different Approaches for Removal of Baseline Wander from ECG Signal, 2nd International Conference and workshop on Emerging Trends in Technology (ICWET)”, Proceedings published by International Journal of Computer Applications® (IJCA).
Akash Kumar Bhoi et al. / Procedia Computer Science 132 (2018) 606–613
Akash Kumar Bhoi / Procedia Computer Science 00 (2018) 000–000
613
[26] S. Azevedo, R.L. Longini, (1980) “Abdominal-lead fetal electrocardiographic R-wave enhancement for heartrate determination”, IEEE Trans. Biomedical Engineering 27(5): 255-260. [27] R.M. Rangayyan, (2001) “Biomedical Signal Analysis: A Case-study Approach”, Wiley–Interscience, New York, 18–28. [28] Yeh Yun-Chi and Wang Wen-June, (2008) “QRS complexes detection for ECG signal: The Difference Operation Method in”. Computer Methods and Programs Biomedicine 91: 245–254. [29] A.K. Bhoi, K.S. Sherpa, B. Khandelwal, (2015) “Classification Probability Analysis for Arrhythmia and Ischemia Using Frequency Domain Features of QRS Complex”, INT. J. BIOAUTOMATION19(4): 531-542 [30] S. Azevedo, R.L. Longini, (1980) “Abdominal-lead fetal electrocardiographic R-wave enhancement for heartrate determination”, IEEE Trans. Biomedical Engineering 27(5): 255-260. [31] R. A. Fisher, (1936) “The Use of Multiple Measurements in Taxonomic Problems”, Annals of Eugenics7: 179–188. [32] R. Duda, P. Hart, D. Stork, (2000) “Pattern Classification”, 2nd ed. New York: Wiley- Interscience. [33] Takeshi Tsutsumi, Yoshiwo Okamoto, Nami Takano, Daisuke Wakatsuki, Takanobu Tomaru, Toshiaki Nakajima (2017) "High -frequency power within the QRS complex in ischemic cardiomyopathy patients with ventricular arrhythmias: Insights from a clinical study and computer simulation of cardiac fibrous tissue", Computers in Biology and Medicine87: 132–140. [34] Takeshi Tsutsumi, Yoshiwo Okamoto, Nami Kubota-Takano, Daisuke Wakatsuki, Hiroshi Suzuki, Kazunori Sezaki, Kuniaki Iwasawa, Toshiaki Nakajima (2014) "Time–frequency analysis of the QRS complex in patients with ischemic cardiomyopathy and myocardial infarction", IJC Heart & Vessels4:177– 187. [35] Feng Jin, Lakshmi Sugavaneswaran, Sridhar Krishnan, Vijay S. Chauhan (2017) "Quantification of fragmented QR S complex using intrinsic time-scale decomposition", Biomedical Signal Processing and Control 31: 513–523.