Heart Sound Detection Paper Published in IEEE T-BME

We are excited to announce that our paper, "A Dual Classifier-Regressor Architecture for Heart Sound Onset/Offset Detection", has been published in IEEE Transactions on Biomedical Engineering (T-BME), one of the top journals in biomedical signal processing.

Led by PhD student Pamuditha Somarathne, this work presents a novel deep learning architecture for precisely detecting the onset and offset of heart sounds — a critical step in automated cardiac auscultation.

The Challenge

Heart sounds, recorded via phonocardiograms (PCG), contain vital diagnostic information. However, accurately pinpointing the exact start and end of each heart sound event (S1 and S2) remains a difficult signal processing challenge, especially in noisy clinical environments. Existing methods often struggle with irregular heart rhythms, murmurs, and background noise.

Our Approach

The paper introduces a dual classifier-regressor architecture that jointly leverages PCG and ECG signals. Rather than treating onset/offset detection as a pure classification or regression problem, this framework combines both:

• Classifier Branch: Identifies the presence and type of heart sound events with high accuracy.
• Regressor Branch: Refines the temporal boundaries of each event to sub-sample precision.

By fusing information from both cardiac signal modalities, the system achieves robust performance across diverse patient populations and recording conditions.

Impact

• Clinical Relevance: Accurate heart sound segmentation is a foundational step for automated diagnosis of heart valve disorders, murmurs, and other cardiac conditions.
• Open Source: The code and trained models are publicly available on GitHub, enabling reproducibility and further research.
• Cross-Signal Fusion: The dual-modality approach demonstrates the value of combining PCG and ECG signals for improved cardiac monitoring.

Congratulations to Pamuditha and the team on this significant publication!

Read more on our Publications page or via DOI.