Elsevier

Expert Systems with Applications

Volume 92, February 2018, Pages 334-349
Expert Systems with Applications

Novel methodology of cardiac health recognition based on ECG signals and evolutionary-neural system

https://doi.org/10.1016/j.eswa.2017.09.022Get rights and content

Highlights

  • New methodology based on single lead and analysis of longer (10-s) ECG signal fragments is proposed.

  • New training based on genetic algorithm coupled with 10-fold cross-validation is employed.

  • 17 classes: normal sinus rhythm + pacemaker rhythm + 15 cardiac disorders are recognized.

  • New feature extraction and selection based on PSD, DFT and GA are employed.

  • Recognition sensitivity at a level of 90.20% (98 errors per 1000 classifications) is promising.

Abstract

This article presents an innovative research methodology that enables the efficient classification of cardiac disorders (17 classes) based on ECG signal analysis and an evolutionary-neural system.

From a social point of view, it is extremely important to prevent heart diseases, which are the most common cause of death worldwide. According to statistical data, 50 million people are at risk for cardiac diseases worldwide. The subject of ECG signal analysis is very popular. However, due to the great difficulty of the task undertaken, and high computational complexity of existing methods, there remains substantial work to perform.

This research collected 1000 fragments of ECG signals from the MIH-BIH Arrhythmia database for one lead, MLII, from 45 patients. An original methodology that consisted of the analysis of longer (10-s) fragments of the ECG signal was used (an average of 13 times less classifications). To enhance the characteristic features of the ECG signal, the spectral power density was estimated (using Welch’s method and a discrete Fourier transform). Genetic optimization of parameters and genetic selection of features were tested. Pre-processing, normalization, feature extraction and selection, cross-validation and machine learning algorithms (SVM, kNN, PNN, and RBFNN) were used.

The best evolutionary-neural system, based on the SVM classifier, obtained a recognition sensitivity of 17 myocardium dysfunctions at a level of 90.20% (98 errors per 1000 classifications, accuracy = 98.85%, specificity = 99.39%, time for classification of one sample = 0.0023 [s]). Against the background of the current scientific literature, these results are some of the best results to date.

Introduction

Diagnosing heart conditions by analyzing ECG signals has been popular for many years and is the basic method used in the prevention of cardiovascular diseases. The wide range of application of ECG signal analysis is due to the fact that it is a simple and non-invasive method that provides substantial valuable information about the function of the circulatory system.

The huge popularity of the ECG signal analysis is also reflected in research. In recent years, the most developed topics related to electrocardiography include: 1) ECG beat detection / classification: (Augustyniak, 2015, Martis, Acharya, Adeli, 2014, Martis, Acharya, Min, 2013, Song, Cho, Kim, Lee, 2015, Yochum, Renaud, Jacquir, 2016), 2) deep learning: (Acharya, Fujita, Lih, Hagiwara, Tan, Adam, 2017, Kiranyaz, Ince, Gabbouj, 2016, Rahhal, Bazi, AlHichri, Alajlan, Melgani, Yager, 2016), 3) principal component analysis: (Castells, Laguna, Sörnmo, Bollmann, Roig, 2007, Ceylan, Ozbay, 2007, Chawla, 2009, Elhaj, Salim, Harris, Swee, Ahmed, 2016, Kallas, Francis, Honeine, Amoud, Richard, 2012, Kanaan, Merheb, Kallas, Francis, Amoud, Honeine, 2011, Kim, Shin, Shin, Lee, 2009, Martis, Acharya, Lim, Suri, 2013, Martis, Acharya, Mandana, Ray, Chakraborty, 2012, Martis, Acharya, Min, 2013, Polat, Gunes, 2007, Rodriguez, Mexicano, Bila, Cervantes, Ponce, 2015, Wang, Chiang, Hsu, Yang, 2013), 4) higher order statistics: (Martis, Acharya, Mandana, Ray, Chakraborty, 2013, Martis, Acharya, Prasad, Chua, Lim, Suri, 2013, Martis, Acharya, Ray, Chakraborty, 2011), 5) feature selection / dimensionality reduction: (Bereta, Burczyński, 2007, Doquire, de Lannoy, Francois, Verleysen, 2011, Doquire, de Lannoy, Francois, Verleysen, 2011, Kishore, Singh, 2015, Lin, Ying, Chen, Lee, 2008b, Llamedo, Martinez, 2011, Mar, Zaunseder, Martineznez, Llamedo, Poll, 2011, Martis, Acharya, Adeli, Prasad, Tan, Chua, et al., 2014, Nasiri, Naghibzadeh, Yazdi, Naghibzadeh, 2009, Oh, Lee, Moon, 2004, Wang, Yang, Teng, Xia, Jensen, 2007, Yeh, Wang, Chiou, 2010, Yu, Lee, 2012, Zhang, Dong, Luo, Choi, Wu, 2014), 6) noise: (Li, Rajagopalan, Clifford, 2014, Pasolli, Melgani, 2015, Roonizi, Sassi, 2016), 7) discrete wavelet transform: (Augustyniak, 2003; Daamouche, Hamami, Alajlan, Melgani, 2012, Elhaj, Salim, Harris, Swee, Ahmed, 2016, Guler, Ubeyli, 2005, Islam, Haque, Tangim, Ahammad, Khondokar, 2012, Kutlu, Kuntalp, 2012, Lin, Du, Chen, 2008a, Martis, Acharya, Min, 2013, Mishra, Thakkar, Modi, Kher, 2012, Thomas, Das, Ari, 2015, Yan, Lu, 2014), 8) independent component analysis: (Chawla, 2009, Elhaj, Salim, Harris, Swee, Ahmed, 2016, Martis, Acharya, Min, 2013, Sarfraz, Khan, Li, 2014, Yu, Chou, 2008, Yu, Chou, 2009), 9) ensemble learning: (Guler, Ubeyli, 2005, Huang, Liu, Zhu, Wang, Hu, 2014, Javadi, Arani, Sajedin, Ebrahimpour, 2013, Mert, Kılıç, Akan, 2012, Osowski, Hoai, Markiewicz, 2004, Osowski, Markiewicz, Hoai, 2008, Sambhu, Umesh, 2013), 10) hybrid systems: (Engin, 2004, Meau, Ibrahim, Narainasamy, Omar, 2006, Osowski, Linh, 2001, Osowski, Markiewicz, Hoai, 2008, Ozbay, Ceylan, Karlik, 2006).

Currently, we observe a very high incidence of cardiovascular disease and the very high mortality caused by them. Despite the preventive measures taken, cardiovascular diseases are the leading cause of death worldwide (17.3 million people per year, accounting for 37% of all deaths (AHA, 2004, AHA, 2016, WHO, 2014)) and the most serious and costly health problems facing the world today (Heron, Smith, 2003, National Center for Health Statistics, 2005). Circulatory system diseases are usually chronic diseases that require long-term and expensive treatment. The tendency for the incidence of cardiovascular diseases will increasingly intensify due to the progressive aging of the population (the number of deaths will increase from 17.3 million in 2016 to 23.6 million in 2030 (AHA, 2004, AHA, 2016, Healthsquare, 2007, WHO, 2014)).

The classification of cardiac disorders based on existing methods based on the calculation of morphological and dynamic features of individual QRS complexes (heart evolution) is difficult and error prone due to the variability of these features in different patients (Padmavathi & Ramakrishna, 2015). For this reason, solutions currently described in the scientific literature do not achieve a satisfactory efficiency (da S. Luz, Schwartz, Cmara-Chvez, & Menotti, 2016).

The existing approaches are also ineffective for certain cardiac disorders, characterized by complex dependencies between subsequent evolutions of heart, for which the most important are “prolapsed” evolutions of heart (time intervals between subsequent heartbeats) and not the QRS complexes that may be correct. The group of these dysfunctions can include pre-excitation syndromes (e.g., Wolff-Parkinson-White syndrome - WPW), atrio-ventricular and atrial-sinus conduction blocks, and elongate PQ intervals.

This is why it is very important to develop specialized software supporting medical diagnostics to more effectively identify heart pathologies earlier and monitor the conditions of patients in real time. The reduction in computational complexity is also an important aspect in the context of deploying the solution in mobile devices.

For recent years, we can distinguish two main approaches in the literature on the automatic recognition of cardiac disorders based on the analysis of ECG signals:

  • classification of QRS complexes (Alvarado, Lakshminarayan, Principe, 2012, de Chazal, O’Dwyer, Reilly, 2004, Mateo, Torres, Aparicio, Santos, 2016, Oster, Behar, Sayadi, Nemati, Johnson, Clifford, 2015, Ye, Kumar, Coimbra, 2012a, Zhang, Luo, 2014),

  • analysis of longer ECG signal fragments (Abawajy, Kelarev, Chowdhury, 2013, Padmavathi, Ramakrishna, 2015, Romero, Serrano, 2001, Vafaie, Ataei, Koofigar, 2014).

It should be noted that the first approach concerning the classification of QRS complexes is substantially more popular. A key element of this approach is the effective detection of QRS complexes. On this basis, it is possible to segment an entire signal into individual QRS complexes and then analyze them using morphological features (determining the shape of the heart evolutions) and dynamic features (determining dependencies between subsequent heart evolutions).

An alternative approach is the analysis of longer, from a single QRS complex (lasting approximately 1 s), signal fragments, usually lasting approximately 10 s; this is the time period corresponding to a standard ECG examination at a cardiologist. Such analysis is based on distinctive feature extraction, for a given disorder, for whole, longer fragments. The identification of heart pathology is based on the extracted features.

Based on a current literature review (Augustyniak, & Tadeusiewicz, da S. Luz, Nunes, de Albuquerque, Papa, Menotti, 2013, da S. Luz, Schwartz, Cmara-Chvez, Menotti, 2016), the typical research methodology in the field of ECG signal analysis consists of

  • 1.

    obtaining data from public databases (MIT-BIH, EDB, AHA, CU, and NST),

  • 2.

    pre-processing and signal normalization,

  • 3.

    QRS detection and ECG signal segmentation,

  • 4.

    extraction of characteristic signal features and rejection of redundant and erroneous information (extraction and selection of features),

  • 5.

    classification of QRS complexes (recognition of heart disorders), e.g., data cross-validation, training, testing and optimization of classifier parameters, and

  • 6.

    evaluation of the obtained results.

In the literature, the most popular method for creating training and test sets is cross-validation, where the two most popular validation schemes are Afkhami, Azarnia, and Tinati (2016) and da S. Luz et al. (2016)

  • class-oriented validation schemes (intra-patient paradigm) - the selection of elements for training and test sets based on signals from the same patient, and

  • subject-oriented validation schemes (inter-patient paradigm) (de Chazal et al., 2004) - the selection of elements for training and test sets based on signals from other patients.

Designing universal algorithms for the general population, not for an individual person, using a subject-oriented validation scheme is a better solution. This solution demonstrates lower effectiveness on the test set but is more reliable and stable and performs better in practice due to the smaller fit of the models to the training set and better knowledge generalization (Afkhami et al., 2016).

The evolutionary-neural system (Rutkowski, 2008) is a hybrid that combines the advantages of two computational intelligence methods: broadly defined Neural Networks (Prieto et al., 2016) and Evolutionary Computation (Back, Hammel, & Schwefel, 1997). With this synergy, we can achieve greater efficiency through better optimization of the classifier tuning by Genetic Algorithm (Holland, 1992) that are parts of the system. In the field of heart disorders recognition, evolutionary-neural systems are also popular and used with success: (Daamouche, Hamami, Alajlan, Melgani, 2012, Dilmac, Korurek, 2015, Ince, Kiranyaz, Gabbouj, 2009, Khazaee, Ebrahimzadeh, 2010, Korurek, Dogan, 2010, Lessmann, Stahlbock, Crone, 2006, Melgani, Bazi, 2008, Shadmand, Mashoufi, 2016).

The main aims of the research were the following:

  • Aim 1 Develop new and effective methods for the automatic recognition of myocardium dysfunctions based on ECG signals modeled on the work of cardiologists.

  • Aim 2 Design algorithms for use in tele-medicine and mobile devices for patient self-control and prevention applications (low computational complexity).

  • Aim 3 Design universal algorithms not for individuals but for the general population.

Based on a literature review (da S. Luz, Nunes, de Albuquerque, Papa, Menotti, 2013, da S. Luz, Schwartz, Cmara-Chvez, Menotti, 2016), it can be stated that the innovative elements of this research include the following:

  • Methodology - a new approach to ECG signal analysis. The designed system is modeled on the work of a cardiologist based on the analysis of longer (10-s) ECG signal fragments, which contain multiple heart evolutions.

  • 17 recognized classes - normal sinus rhythm + pacemaker rhythm + 15 cardiac disorders.

  • Genetic training and optimization of classifiers - a genetic algorithm coupled with 10-fold cross-validation for signal feature selection and classifier parameter optimization.

The innovative elements of this research in the analysis of long (10-s) fragments of ECG signals include the following:

  • Feature extraction - strengthen the characteristic features of signals by estimating the power spectral density using the Welch method and the discrete Fourier transform (data analysis in the frequency domain for several Hamming window widths).

  • Genetic selection of features – the elimination of redundant features (frequency components of the power spectral density of the ECG signal) by a genetic algorithm.

Section snippets

ECG database

For research purposes, the ECG signals were obtained from the http://www.physionet.org PhysioNet (Goldberger et al., 2000) service from the MIT-BIH Arrhythmia (Moody & Mark, 2001) database. The created database with ECG signals is described below.

  • The ECG signals were from 45 patients.

  • The ECG signals contained 17 classes: normal sinus rhythm, pacemaker rhythm, and 15 types of cardiac dysfunctions (for each of which at least 10 signal fragments were collected).

  • All ECG signals were recorded at a

Results

All combinations of methods (2.4) have been tested on a smaller database containing 744 ECG fragments and 29 patients. This article presents the results for the 8 paths that have achieved the highest sensitivity (SEN): 1 type of normalization x 2 Hamming window widths x 4 types of classifiers. The results obtained were very similar for both variants of databases (744 and 1000 ECG fragments).

The study utilized the MATLAB R2014b environment together with the LIBSVM library (Chang & Lin, 2011).

Hypothesis

The results obtained in all experiments confirmed the thesis: the application of the proposed methodology will enable the automatic, efficient, universal, low computational complexity and fast recognition of heart disorders based on ECG signal analysis and the evolutionary-neural system.

The confirmation of this statement is given by the obtained results, summarized in Table 3, Table 4, Table 5 and in 4.11 Computational complexity, 4.9 Times. The presented results show that the recognition

Conclusion

The aim of the conducted research was to develop a new methodology that enables the efficient recognition of myocardium dysfunctions (17 classes: normal sinus rhythm + pacemaker rhythm + 15 heart disorders), based on analysis of 10-s fragments of ECG signals and an evolutionary-neural system. In this research, 1000 fragments of ECG signals were analyzed from the MIH-BIH Arrhythmia database for one lead, MLII, from 45 patients. Four experiments were conducted, during which many methods were

Acknowledgments

We thanks Professor R. Tadeusiewicz and Professor Z. Tabor for their valuable advice and guidance on the preparation of this paper.

P. Pławiak was born in Ostrowiec, Poland, in 1984. He obtained his M.Sc. degree in Electronics and Telecommunications and his Ph.D degree with honors in Biocybernetics and Biomedical Engineering at the AGH University of Science and Technology, Cracow, Poland, in 2012 and 2016, respectively. He is an Assistant Professor with the Institute of Telecomputing, Cracow University of Technology, Cracow, Poland. His research interests include machine learning algorithms and computational intelligence

References (118)

  • L.I. Kuncheva

    Combining pattern classifiers: Methods and algorithms

    (2004)
  • Q. Li et al.

    A machine learning approach to multi-level ECG signal quality classification

    Computer Methods and Programs in Biomedicine

    (2014)
  • T. Mar et al.

    Optimization of ECG classification by means of feature selection

    IEEE Transactions on Biomedical Engineering

    (2011)
  • R.J. Martis et al.

    Computer aided diagnosis of atrial arrhythmia using dimensionality reduction methods on transform domain representation

    Biomedical Signal Processing and Control

    (2014)
  • R.J. Martis et al.

    Application of principal component analysis to ecg signals for automated diagnosis of cardiac health

    Expert Systems with Applications

    (2012)
  • R.J. Martis et al.

    ECG Beat classification using PCA, LDA, ICA and discrete wavelet transform

    Biomedical Signal Processing and Control

    (2013)
  • R.J. Martis et al.

    Application of higher order statistics for atrial arrhythmia classification

    Biomedical Signal Processing and Control

    (2013)
  • Y.P. Meau et al.

    Intelligent classification of electrocardiogram (ECG) signal using extended Kalman filter (EKF) based neuro fuzzy system

    Computer Methods and Programs in Biomedicine

    (2006)
  • F. Melgani et al.

    Classification of electrocardiogram signals with support vector machines and particle swarm optimization

    IEEE Transactions on Information Technology in Biomedicine

    (2008)
  • A. Mert et al.

    Evaluation of bagging ensemble method with time-domain feature extraction for diagnosing of arrhythmia beats

    Neural Computing and Applications

    (2012)
  • A. Mishra et al.

    Comparative analysis of wavelet basis functions for ECG signal compression through compressive sensing

    International Journal of Computer Science and Telecommunications

    (2012)
  • G.B. Moody et al.

    The impact of the mit-bih arrhythmia database

    IEEE Engineering in Medicine and Biology Magazine

    (2001)
  • J.A. Nasiri et al.

    ECG arrhythmia classification with support vector machines and genetic algorithm

    Computer modeling and simulation, 2009. ems ’09. third uksim european symposium on

    (2009)
  • National Center for Health Statistics (Ed.)

    Health, United States, 2005 with chartbook on the health of Americans

  • S. Osowski et al.

    ECG beat recognition using fuzzy hybrid neural network

    IEEE Transactions on Biomedical Engineering

    (2001)
  • R. Rodriguez et al.

    Feature extraction of electrocardiogram signals by applying adaptive threshold and principal component analysis

    Journal of Applied Research and Technology

    (2015)
  • E.K. Roonizi et al.

    A signal decomposition model-based bayesian framework for ECG components separation

    IEEE Transactions on Signal Processing

    (2016)
  • L. Rutkowski

    Computational intelligence: Methods and techniques

    (2008)
  • B. Scholkopf et al.

    Learning with kernels: Support vector machines, regularization, optimization, and beyond

    (2001)
  • S. Shadmand et al.

    A new personalized ECG signal classification algorithm using block-based neural network and particle swarm optimization

    Biomedical Signal Processing and Control

    (2016)
  • S. Smith

    Digital signal processing: A practical guide for engineers and scientists

    (2002)
  • M. Sokolova et al.

    A systematic analysis of performance measures for classification tasks

    Information Processing & Management

    (2009)
  • M.-H. Song et al.

    New real-time heartbeat detection method using the angle of a single-lead electrocardiogram

    Computers in Biology and Medicine

    (2015)
  • R.G. Afkhami et al.

    Cardiac arrhythmia classification using statistical and mixture modeling features of ECG signals

    Pattern Recognition Letters

    (2016)
  • AHA

    International cardiovascular disease statistics

    (2004)
  • AHA

    Heart disease, stroke and research statistics at-a-glance

    (2016)
  • E. Alpaydin

    Introduction to machine learning

    (2014)
  • N.S. Altman

    An introduction to kernel and nearest-neighbor nonparametric regression

    The American Statistician

    (1992)
  • A.S. Alvarado et al.

    Time-based compression and classification of heartbeats

    IEEE Transactions on Biomedical Engineering

    (2012)
  • P. Augustyniak

    A robust heartbeat detector not depending on ECG sampling rate

    Engineering in medicine and biology society (embc), 2015 37th annual international conference of the ieee

    (2015)
  • Augustyniak, P., & Tadeusiewicz, R. (2009). Ubiquitous cardiology - emerging wireless telemedical application., pp....
  • T. Back et al.

    Evolutionary computation: Comments on the history and current state

    IEEE Transactions on Evolutionary Computation

    (1997)
  • Y. Bazi et al.

    Domain adaptation methods for ECG classification

    Computer medical applications (iccma), 2013 international conference on

    (2013)
  • M. Bereta et al.

    Comparing binary and real-valued coding in hybrid immune algorithm for feature selection and classification of ECG signals

    Engineering Applications of Artificial Intelligence

    (2007)
  • J. Bergstra et al.

    Random search for hyper-parameter optimization

    Journal of Machine Learning Research

    (2012)
  • D.S. Broomhead et al.

    Radial basis functions, multi-variable functional interpolation and adaptive networks

    Complex Systems 2

    (1988)
  • F. Castells et al.

    Principal component analysis in ecg signal processing

    EURASIP Journal of Applied Signal Processing

    (2007)
  • R. Ceylan et al.

    Comparison of fcm, PCA and WT techniques for classification ECG arrhythmias using artificial neural network

    Expert Systems with Applications

    (2007)
  • C.-C. Chang et al.

    LIBSVM: A library for support vector machines

    ACM Transactions on Intelligent Systems and Technology

    (2011)
  • M. Chawla

    A comparative analysis of principal component and independent component techniques for electrocardiograms

    Neural Computing and Applications

    (2009)
  • Cited by (216)

    View all citing articles on Scopus

    P. Pławiak was born in Ostrowiec, Poland, in 1984. He obtained his M.Sc. degree in Electronics and Telecommunications and his Ph.D degree with honors in Biocybernetics and Biomedical Engineering at the AGH University of Science and Technology, Cracow, Poland, in 2012 and 2016, respectively. He is an Assistant Professor with the Institute of Telecomputing, Cracow University of Technology, Cracow, Poland. His research interests include machine learning algorithms and computational intelligence methods (e.g., artificial neural networks, genetic algorithms, fuzzy systems, support vector machines, k-nearest neighbors, and hybrid systems), pattern recognition, signal processing and analysis, data analysis and data mining, sensor techniques, medicine, biocybernetics, and biomedical engineering.

    View full text