TriFusion-PCGNet: A Hybrid Multi-Branch Deep Learning Network for Energy-Aware Normal/Abnormal Heart Sound Classification

Background/Objectives: Cardiovascular diseases remain a leading cause of death worldwide, and early diagnosis is essential. Phonocardiogram (PCG) signals provide a non-invasive and inexpensive means of cardiac assessment. However, accurate interpretation of PCG signals requires clinical expertise. In this paper, we propose TriFusion-PCGNet, a hybrid multi-branch deep learning framework for automated heart sound classification. Methods: The proposed method was evaluated on the PhysioNet/Computing in Cardiology Challenge 2016 dataset. Three complementary PCG representations—the raw signal, band-pass filtered signal, and Hilbert envelope signal—were processed through dedicated feature extraction branches using residual convolutional networks, dilated residual convolutions, and a BiGRU with a temporal attention mechanism. The extracted features were fused and classified using fully connected layers. Performance was evaluated using the original train–validation split and stratified 10-fold cross-validation. Results: Under a strictly separated record-level 10-fold cross-validation protocol, TriFusion-PCGNet achieved 90.87% mean accuracy and 95.34% AUC, with outer test folds isolated from training and model selection. Subject-level analysis showed substantially lower performance, emphasizing the importance of subject-disjoint evaluation for patient-level generalization. Conclusions: TriFusion-PCGNet integrates complementary PCG representations for automated normal/abnormal heart sound classification. However, the substantially lower subject-disjoint performance indicates that the record-level results should not be interpreted as evidence of patient-independent generalization, and further validation on independent subject-annotated datasets is required.