A Dual Classifier-Regressor Architecture for Heart Sound Onset/Offset Detection

Area of Science:

• Cardiology
• Biomedical Engineering
• Signal Processing

Background:

• Accurate identification of heart sounds (S1 and S2) from phonocardiogram (PCG) signals is crucial for diagnosing various cardiac conditions.
• Deep learning models, particularly those inspired by image segmentation and aided by synchronized electrocardiograms (ECG), have shown potential for PCG analysis.
• Existing methods often focus on point-wise segmentation, but identifying the precise onset and offset of heart sounds remains a challenge.

Purpose of the Study:

• To develop and evaluate a novel method for identifying the onset and offset of the first (S1) and second (S2) heart sounds in PCG signals.
• To leverage synchronized ECG signals and their keypoints to enhance the accuracy of heart sound detection.
• To shift the focus from point-wise segmentation to the precise localization of heart sound transitions.

Main Methods:

• A joint classifier-regressor architecture was proposed to predict both the probability and precise location of S1 and S2 sound onsets and offsets.
• The method incorporates synchronized ECG signals and their derived keypoints to improve the detection of heart sounds within the PCG signal.
• The model was trained and evaluated on the PhysioNet/CinC 2016 dataset, the largest publicly available dataset for this task.

Main Results:

• The proposed approach achieved state-of-the-art performance on the PhysioNet/CinC 2016 dataset.
• Sensitivity of 0.97 and a positive predictive value of 0.98 were obtained for identifying the midpoints of S1 and S2 segments.
• The method accurately identified the onset/offset locations of heart sounds with a mean error of 11.11 ms.

Conclusions:

• Identifying the transitions (onset/offset) of heart sounds simplifies the analysis, leading to improved model training and inference.
• The developed method offers a significant advancement in the automated analysis of PCG signals for cardiac diagnostics.
• The approach has potential for adaptation to locate regions of interest in other physiological signals like respiration or blood pressure.