DPPAT: Dual-Level Periodic Pattern-Aware Transformer for Heart Sound Murmur Identification.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiology

Background:

• Early heart disease screening relies on accurate heart sound classification for murmur identification.
• Identifying murmurs is challenging due to signal weakness and noise interference.
• Current methods underutilize periodic patterns inherent in heart sounds.

Purpose of the Study:

• To propose a novel Dual-level Periodic Pattern-Aware Transformer (DPPAT) for enhanced murmur identification.
• To implicitly leverage periodic priors in heart sound signals without requiring cycle segmentation.
• To improve the accuracy and generalizability of heart murmur detection algorithms.

Main Methods:

• Developed a Dual-level Periodic Pattern-Aware Transformer (DPPAT) model.
• Implemented an Adaptive Period-Aligned Window Selection algorithm for regional-level feature extraction.
• Utilized a Periodic Pattern Attention module to suppress noise and extract periodic components.
• Integrated periodic features at the global-level for enhanced murmur discriminative feature identification.

Main Results:

• Achieved a weighted accuracy of 84.27% and an F1-score of 70.38% on the 2022 George B. Moody PhysioNet Challenge dataset.
• Demonstrated generalizability on two additional public datasets containing heart and respiratory sounds.
• Attention visualizations confirmed the model's focus and decision-making basis for murmur identification.

Conclusions:

• The DPPAT model effectively leverages periodic priors for improved heart murmur identification.
• The proposed method offers a promising advancement in automated cardiac auscultation and early disease screening.
• DPPAT shows strong performance and generalizability across different datasets, highlighting its clinical potential.