Feature-Based Cardiac Cycle Segmentation in Phonocardiogram Recordings

Abstract

Phonocardiogram (PCG) conveys crucial information for cardiac health evaluation in ambulatory care and is an essential diagnostic test for heart assessment. Thus, identification and positioning of the first and second heart sound within PCG is a vital step in automatic heart sound analysis. This study proposes a solution for individual cardiac cycle segmentation of PCG recordings. It extracts a rich set of features that are used for the segmentation of each cardiac cycle in a PCG recording by localizing the PCG peaks, S1 and S2. To accomplish this objective, a rich set of 66 discriminative features are selected and extracted from each frame in a PCG recording and several classifiers are evaluated to find out the one that achieves the highest segmentation accuracy. Finally, a post-processing method is proposed to reduce the classification noise and hence improve the segmentation performance Contrary to the earlier methods proposed in the literature, this method is evaluated on one of the largest datasets available consisting of 48 877s PCG recordings. The proposed method has achieved F1-score of 93.45%, and Sensitivity and Specificity values of 94.23% and 98.16% respectively. Moreover, it has been tested on the Pascal benchmark dataset, and has achieved Sensitivity and Specificity values of 96.42% and 98.12%, respectively.