ECG heartbeat classification by means of variable rational projection
Graphical abstract
Introduction
Electrocardiogram (ECG) is the most widely used non-invasive diagnostic tool to monitor cardiac activity. Computer-assisted analysis has been and still is a relevant research topic. In practice, reliable automatic cardiac disorder detection can be of significant help in managing certain clinical situations, or long-term monitoring. In the latter case, the human-based interpretation is time-consuming and resource intensive, yet the timely diagnosis is an important issue.
In this paper, we focus on the detection of specific arrhythmia types. Cardiac arrhythmia is a group of conditions where the heart shows abnormal activity or behavior. These conditions may or may not be directly life-threatening, but still need medical therapy to avoid complications. Several types of arrhythmia have been investigated so far using computer-assisted techniques. One usual way is the classification of heartbeats into predefined classes. The ANSI/AAMI EC57:1998–2012 standards [1] recommend 5 (one normal and four abnormal) arrhythmia superclasses, and in a more refined categorization 15 classes. PhysioNet [2] provides 16 classes, i.e. a paced beat class is added to the previous 15 classes.
In recent years, several methods were published on automatic heartbeat classification [3], [4], [5], [6], [7], [8], [9], [10], [11], in which a proper feature extraction method along with a classifier were constructed. The purpose of the feature extraction methods is to find a characterization of heartbeats based on mathematical and physiological models. The clinical practice show that arrhythmia can cause both heart rhythm irregularities and waveform shape distortions. Therefore a common approach is to catch these dynamic (heart rhythm related) and morphological (waveform related) differences around the heartbeat, and then utilize a combination of dynamic and morphological features. For dynamic descriptors, RR interval features are the most widely employed. In early works two features, the previous and post-RR intervals, were used. Later two more dynamic features, called local and average RR features proposed by de Chazal et al. [6] and later revised by Ye et al. [9], were added. The first is reflecting to short and the second to long-term behavior of heart rhythm. In the work of Lin et al. [10], the RR interval features are normalized in order to reduce the effect of different heart rates in different records. Morphological features are generated by appropriate transformations. A great variety of them can be found in previous works. They include Hermite coefficients [3], [5], high-order statistics (HOS) features [5], waveform shape features [7], [12], wavelet features [4], [8], [9], independent component analysis (ICA) [8], [9], etc. Then several machine learning algorithms have were applied for classification, such as artificial neural network (ANN) [4], conditional random field (CRF), decision tree (DT) [7], dynamic Bayesian network (DBN), linear discriminants (LD), self-organizing map (SOM) [3], and support vector machine (SVM) [5], [8], [9]. For a review on the relevant literature, we refer the reader to Sansone et al. [13] and Luz et al. [14].
The standard test material in heartbeat classification is the MIT-BIH Arrhythmia Database [15]. It is widely used as benchmark database for the validation of classification methods. The database contains 48 half-hour two-lead ambulatory ECG recordings, obtained from 47 subjects. It consists of more than 100,000 heartbeats for which reference annotations are provided by cardiologists. They include the location of the R peaks and labels referring to one of the 16 arrhythmia classes. The importance of the database is twofold. It provides a comprehensive, annotated test material for validation and performance evaluation, and allows direct comparison of various methods. The database and related materials are available on PhysioNet [2].
Based on the adopted evaluation scheme, the reference works can be distinguished as ‘class-oriented’ and ‘subject-oriented’ (see e.g. Escalona et al. [11]). In this paper we present an effective heartbeat classification method evaluated against both strategies. It is based on a variable projection [16] transform involving rational functions. The idea of representation of ECG signals with rational functions goes back to Fridli et al. [17], [18]. Since then these systems have been applied also for approximation, compression of ECG signals [19], [20], [21], [22], [23], for heartbeat detection [24], and also for detecting epileptic seizures in EEG signals [25], [26]. The problem of classification, however, needs an approach different from those in the previous papers. There are several reasons that explain why the rational function systems perform well in ECG processing, and why we choose it for heartbeat classification. Here we only mention adaptivity and flexibility.
We take a patient-specific representation of the heartbeats using adaptive rational transforms. It involves the identification of the optimal system parameters. The coefficients of the projection and also the system parameters themselves are used as morphological features. They will be extended with dynamic features derived from the RR intervals. Then a support vector machine (SVM) classifier is applied followed by two-lead fusion. Our method is evaluated on the MIT-BIH Arrhythmia Database with 16 or 5 classes, adopting the ‘class-oriented’ and ‘subject-oriented’ scheme, following the general methodology described in de Chazal et al. [6] and Luz et al. [14]. We note that this work relies on our previous one [27], but significantly improves it in several respects. Namely, a more stable and reliable rational heartbeat model and parameter optimization method is introduced, according to an extensive study of different possibilities. Moreover the system parameters are built in the feature vector. Also the segmentation, and the classification process are fine-tuned.
We built a MATLAB framework for the implementation and performance evaluation tests. The data downloaded from PhysioNet were converted using the WaveForm DataBase (WFDB) Toolbox [28], the rational transform routines are based on the Rational Approximation and Interpolation Toolbox (RAIT) [29], and the LIBSVM package [30] was adapted for the SVM-related methods. The codes are available at http://bognargergo.web.elte.hu/ecg/, and so our results are reproducible.
Section snippets
Variable rational orthogonal projection and heartbeat representation
Since we will use rational projections to model the ECG signals here we provide a short summary about them. For details we refer the reader to Heuberger et al. [31].
Results and discussion
In this section we present our results, including comparison to the state-of-the-art methods, and robustness tests. Moreover, we point out the progress made since our preliminary work [27].
Conclusion
In this paper we presented a new ECG classification algorithm. The novelty in our approach was the application of a variable projection method employing Malmquist–Takenaka systems of rational functions. The consequence of the adaptivity of our method is that both the system parameters and the coefficients of the projections carry medical information. In the optimization process we balanced the conditions on approximation and representation. Patient specific morphological descriptors were
Authors’ contributions
G. Bognár: Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data Curation, Writing – Original Draft, Visualization
S. Fridli: Conceptualization, Methodology, Formal analysis, Writing – Review & Editing, Supervision, Project administration, Funding acquisition.
Acknowledgement
EFOP-3.6.3-VEKOP-16-2017-00001: Talent Management in Autonomous Vehicle Control Technologies – The Project is supported by the Hungarian Government and co-financed by the European Social Fund. This research was supported by the Hungarian Scientific Research Funds (OTKA) No K115804.
Conflicts of interest: The authors declare no conflicts of interest.
References (39)
- et al.
ECG-based heartbeat classification for arrhythmia detection: a survey
Comput. Methods Programs Biomed.
(2016) - et al.
Epileptic seizure detection in long-term EEG records using sparse rational decomposition and local Gabor binary patterns feature extraction
Knowl.-Based Syst.
(2017) Testing and Reporting Performance Results of Cardiac Rhythm and ST Segment Measurement Algorithms, Standard ANSI/AAMI EC57:2012
(2012)PhysioBank, physioToolkit, and physioNet: components of a new research resource for complex physiologic signals
Circulation
(2000)- et al.
Clustering ECG complexes using Hermite functions and self-organizing maps
IEEE Trans. Biomed. Eng.
(2000) - et al.
Classification of ECG arrhythmias using multi-resolution analysis and neural networks
Proc. Conf. Convergent Technol. Asia-Pacific Region, vol. 1
(2003) - et al.
Support vector machine-based expert system for reliable heartbeat recognition
IEEE Trans. Biomed. Eng.
(2004) - et al.
Automatic classification of heartbeats using ECG morphology and heartbeat interval features
IEEE Trans. Biomed. Eng.
(2004) - et al.
Real-time classification of ECGs on a PDA
IEEE Trans. Inf. Tech. Biomed.
(2005) - et al.
ECG arrhythmias recognition system based on independent component analysis feature extraction
Proc. IEEE Region 10 Conf.
(2006)
Heartbeat classification using morphological and dynamic features of ECG signals
IEEE Trans. Biomed. Eng.
Heartbeat classification using normalized RR intervals and morphological features
Math. Problem Eng.
Electrocardiogram classification using reservoir computing with logistic regression
IEEE J. Biomed. Health Inform.
Heartbeat classification using feature selection driven by database generalization criteria
IEEE Trans. Biomed. Eng.
Electrocardiogram pattern recognition and analysis based on artificial neural networks and support vector machines: a review
J. Healthc. Eng.
The impact of the MIT-BIH arrhythmia database
IEEE Eng. Med. Biol. Mag.
The differentiation of pseudo-inverses and nonlinear least squares problems whose variables separate
SIAM J. Numer. Anal.
Rational function systems in ECG processing
Biorthogonal systems to rational functions
Ann. Univ. Sci. Budapest. Sect. Comp.
Cited by (8)
ECG-based heartbeat classification using exponential-political optimizer trained deep learning for arrhythmia detection
2023, Biomedical Signal Processing and ControlExploring deep features and ECG attributes to detect cardiac rhythm classes
2021, Knowledge-Based SystemsCitation Excerpt :CNN [35] and DNN [15] deep learning structures, as well as hybrid models such as CNN-LSTM [34], CNN-spiking neural network (SNN) [38], denoising autoencoder-1D CNN [52], and the Stockwell transform-based two-dimensional residual network [17] have been used for arrhythmia classification studies. A two-dimensional residual network model structure based on the Stockwell transformation attained the highest accuracy (99.73%) [17] among the studies examined. Most of these works had been done using public databases.
Chronological golden search optimization-based deep learning for classification of heartbeat using ECG signal
2023, Computer Methods in Biomechanics and Biomedical Engineering: Imaging and VisualizationTwo-Stage Classifier for Resource Constrained On-Board Cardiac Arrhythmia Detection
2023, IEEE Transactions on Instrumentation and Measurement