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

Overview

This project focused on improving automated heart sound analysis by accurately detecting the onset and offset of key cardiac sounds (S1 and S2) from phonocardiogram (PCG) and electrocardiogram (ECG) signals.

Traditional methods segment heart sounds using point-wise classification, which is often sensitive to noise and temporal variability. In contrast, we introduce a transition-focused approach that detects physiologically meaningful boundaries in the signal instead of classifying every time point. By combining synchronised ECG signals with PCG data, the system achieves highly precise segmentation while improving robustness in real-world and noisy conditions.

The project aims to make cardiac signal analysis more reliable and easier to deploy in clinical and wearable sensing scenarios.

Vision

Accurate heart sound segmentation is a critical step toward automated diagnosis of cardiovascular conditions such as valve disorders, arrhythmias, and heart failure. However, many existing approaches rely on complex post-processing pipelines or struggle under real-world noise.

This project shifts the focus from dense signal classification to key transition detection, simplifying the learning problem while aligning with the physiological structure of cardiac cycles. The vision is to create models that are:

• Robust to signal variability and noise
• Physiologically interpretable
• Easier to generalize across datasets and devices
• Adaptable to other physiological signals beyond heart sounds

Example Applications

The proposed framework enables a range of future research and applied scenarios:

• Automated Cardiac Analysis: Reliable identification of heart sound boundaries for downstream diagnostic tools.
• Wearable Health Monitoring: Integration into portable ECG-PCG sensing systems for continuous monitoring.
• Clinical Decision Support: Improving consistency in heart sound interpretation compared to manual auscultation.
• General Physiological Segmentation: Extension of transition-detection techniques to respiration, blood pressure, or muscle activity signals.

Evaluation

The system was evaluated on the publicly available PhysioNet/CinC 2016 heart sound dataset and demonstrated state-of-the-art performance.

Key outcomes include:

• High sensitivity and positive predictive value in detecting heart sound midpoints
• Average onset/offset localization error of approximately 11 ms
• Improved robustness through a joint classifier–regressor architecture
• Effective use of ECG cues to improve temporal accuracy

Experiments showed that predicting transition events simplifies learning and improves consistency compared to conventional point-wise segmentation methods.

Open Source & Availability

The project provides an open research implementation and resources for reproducibility.

Code repository: https://github.com/aid-lab-org/PCG-ECG-Segmentation.git