ECG analysis for arrhythmia detection and classification

Abstract

Though various techniques have been suggested for the analysis of ECG signals, interpretation of these signals, especially as they affect human health, has posed some difficulties. Consequently, the best way of interpreting these physiological signals by electric measurements from the body surface in terms of cardiac electric activity remains an active research topic till today. This research tackles three problems related to ECG analysis namely, parametric modeling, period normalization (interpolation) and classification of arrhythmias. In order to model the signal, each heartbeat is first mapped into a new domain where the transform coefficients vector would be sparse. The coefficients vector is then approximated to a sum of damped sinusoids. The transform matrix is generated based on the combination of linear prediction (LP) and the singular values decomposition (SVD) of the LPC filter impulse response matrix. This approach leads to relatively satisfactory compression ratio (CR) as compared to existing techniques. Though parametric modeling of ECG signals has a central role in real time transmission and classification of heart abnormalities (arrhythmias), the compression ratios achieved are not suitable for storage purpose. Therefore, 2D ECG compression schemes are adopted where the beats of differing periods should be equalized to the same period length and then arranged in an image matrix before the application of image compression algorithm. Limitations of the existing techniques for ECG period equalization are highlighted and a new frequency domain approach for period normalization has been developed. The proposed approach is signal dependent and able to adapt to the signal characteristics. An analytical model to generate basis functions has also been developed. The merits of the proposed technique are appreciated when compared to other techniques commonly used in the literature. Finally, an algorithm for arrhythmia classification that conforms to the recommended practice of the Association for the Advancement of Medical Instrumentation (AAMI) is presented. Three inter-patient classification scenarios have been considered namely, detection of ventricular ectopic beats (VEBs), detection of supraventricular ectopic beats (SVEBs) and the multiclass recommended taxonomy.A novel set of features extraction via the application of orthogonal transformation of the ECG signal has been developed. These features in conjunction with some commonly used features are fed into the Regularized Least Squares Classifier (RLSC) with linear kernel. The proposed classification scheme shows good separation capability between the classes of ECG arrhythmias as it has achieved a Balanced Classification Rate (BCR) of 83.9 % for the multiclass scenario which is comparable to the state-of-the-art performance of automatic arrhythmia classification algorithms.