Arrhythmia detection using signal-adapted wavelet preprocessing for support vector machines

Rate-based arrhythmia recognition algorithms in implantable cardioverter-defibrillators (ICDs) are of limited reliability in some clinical situations. In such cases, the inclusion of the morphological features of endocardial electrograms can improve the performance. In this study, we present a coupled signal-adapted wavelet and support vector machine (SVM) arrhythmia detection scheme. Within the scope of an electrophysiological examination, data segments were recorded during normal sinus rhythm (NSR) and ventricular tachycardia (VT). Consecutive beats were selected as morphological activation patterns of NSR and VT. These patterns were represented by their multi-level concentrations. For this, a signal-adapted and highly efficient lattice structure-based wavelet decomposition technique was employed which maximizes the class separability and takes into account the final classification of NSR and VT by SVMs with radial, compactly-supported kernels. In an automated analysis of an independent test set, our hybrid scheme outperformed other methods and classified all patterns correctly without overlap.