A dual classifier-regressor architecture for heart sound onset/offset detection

Objective: Identifying the first (S1) and second (S2) heart sounds from phonocardiogram (PCG) signals is an essential step in automating the diagnosis of cardiac conditions such as irregular heartbeat, valve misfunctions, and heart failure. Recent research inspired by image segmentation has shown promise in utilising deep neural networks for point-wise PCG segmentation with the support of synchronised electrocardiograms (ECG). This paper shifts the focus from point-wise segmentation to identifying the onset/offset of S1 and S2 in the PCG signal. Methods: We incorporate the ECG signal and its keypoints to improve the detection of the heart sounds. Our proposed method employs a joint classifier-regressor architecture for predicting the probability and the location of onset/offset in the PCG. Results: When evaluated on the largest publicly available PhysioNet/CinC 2016 dataset, the proposed approach outperforms existing state-of-the-art methods, achieving a sensitivity of 0.97 and a positive predictive value of 0.98 in identifying midpoints of S1 and S2 segments. It also identifies the onset/offset locations with an 11.11 ms error. Conclusion: It is evident that identifying the transitions simplifies, leading to better training and inference. Significance: In addition to achieving state-of-the-art results, this proposed approach could also be adapted for locating regions of interest in other physiological signals, such as respiration, blood pressure, or muscle activity.