A switching feature extraction system for ECG heartbeat classification

This study compared two methods for extracting ECG waveshape features useful for heartbeat classification. The first method (segmented waveshape features) sampled the ECG waveshape between the P and T waveform boundaries, calculated QRS and T wave durations, and used RR-interval features. The second method (fixed interval waveshape features) used a fixed window to capture the ECG waveshape and RR-interval features. Data were obtained from the MIT-BIH arrhythmia database. We investigated the problem of discriminating between normal, supraventricular (SVEB), ventricular (VEB), fusion and unknown beat classes. When the P, QRS and T wave boundaries could be found reliably, the segmented waveshape features resulted in more balanced performance for discriminating SVEB and VEB beats than the fixed interval waveshape features. A hybrid approach using segmented and fixed interval features when waveform boundaries could be reliably found, and fixed interval features otherwise was the most robust solution. Using the AAMI recommendations for cardiac rhythm disturbances the hybrid approach resulted in a sensitivity of 69%, a positive predictivity of 31% and a false positive ratio (FPR) of 6.6% for SVEB class. For the VEB class the sensitivity was 80%, the positivity predictivity was 85% and the FPR was 1.0%.