Feasibility of wavelet-kernel novelty detectors for identifying ventricular tachycardia

Hybrid wavelet-support vector classifiers have recently been suggested as efficient rate independent recognition scheme for endocardial electrograms (EEs) usable in implantable cardioverter defibrillators (ICDs). However, these schemes involve waveforms of the normal sinus rhythm (NSR) as well as the pathological rhythm. In this study, we assess the feasibility of hybrid wavelet-kernel novelty detectors to discriminate between NSR and ventricular tachycardia (VT). These schemes only involve waveforms of NSR for the learning task. Consecutive beats were selected as morphological patterns of NSR and VT. These patterns were represented by their multilevel concentrations using a lattice structure based wavelet decomposition. The lattice angles were optimized for kernel based novelty detectors. In a blind analysis of the independent test set, our scheme outperformed the correlation waveform analysis with best fit alignment which is a well accepted method for arrhythmia recognition. However, the performance of wavelet-support vector classifiers was not achievable. We conclude that wavelet-kernel novelty detectors are feasible for the rate independent identification of VTs. Our scheme is of a low complexity and suitable for efficient real-time implementations. However, further studies are needed to evaluate their performance in the clinical practice.