A review of kernel methods in ECG signal classification

Abstract

Kernel methods have been shown to be effective in the analysis of electrocardiogram (ECG) signals. These techniques provide a consistent and well-founded theoretical framework for developing nonlinear algorithms. Kernel methods exhibit useful properties when applied to challenging design scenarios, such as: (1) when dealing with low number of (potentially high dimensional) training samples; (2) in the presence of heterogenous multimodalities; and (3) with different noise sources in the data. These characteristics are particularly appropriate for biomedical signal processing and analysis, and hence, the widespread of these techniques in biomedical signal processing in general, and in ECG data analysis in particular. Specifically, kernel methods have improved the performance of both parametric linear methods and neural networks in applications such as cardiac beat detection in 12-lead ECG, detection of electrocardiographic changes in partial epileptic patients, automatic identification of reliable heart rates, detection of obstructive sleep apnea, automatic seizure detection in the newborn, or cardiac sound murmurs classification, just to name a few. This chapter provides a survey of applications of kernel methods in the context of ECG signal analysis. The chapter summarizes the theory of kernel methods, and studies the different application domains. Noting that the vast majority of applications in the literature reduce to the use of the standard support vector machine, we pay special attention to other kernel machines available in the literature that may be of interest for practitioners. Finally, we foresee some other future research lines in the development of specific kernel methods to deal with the data peculiarities.