A Fine-Grained Sub-Peak Detection Framework for Bone-conducted Phonocardiograms via Residual U-Net

Shengjie Zhang1, Yumin Li1, Li Ling1, Junjie Pan2, Chaohong Liu2, Huan Li2, Yanan Zhou2, Chenghao Sui2, Wenkai Wang3, Jianqing Li1, Chengyu Liu1
1Southeast University, 2Goertek Technology Co., 3Goertek


Abstract

Fine-grained localization of the first (S1) and second (S2) heart sound sub-peaks is important for cardiac mechanics evaluation, yet challenging in wearable bone-conducted phonocardiogram (BC-PCG) signals. To localize five prominent sub-peaks using multi-modal signals, an improved one-dimensional Residual U-Net is proposed, featuring a sparsity-aware composite heatmap loss to address extreme feature sparsity. A post-processing pipeline using non-maximum suppression (NMS) and conditional random field (CRF) Viterbi decoding enforces physiological temporal constraints. Under leave-one-subject-out cross-validation (LOSOCV) on 12 subjects, the optimal PCG and electrocardiogram (ECG) configuration achieved an 86.57% macro-average F1-score and a 19.36 dB signal-to-noise ratio (SNR), effectively circumventing dynamic temporal misalignments introduced by photoplethysmogram (PPG) integration to ensure robust localization despite physiological intra-wave imbalances. This framework enhances BC-PCG interpretability and facilitates cardiovascular disease diagnostics.