Lung-Sound Respiratory Disease Classification via Multiple-Instance Learning

We propose a multi-channel multiple–instance learning (MIL) framework for automated respiratory disease classification that addresses acoustic heterogeneity among recording site conditions, handles weak supervision, and reflects spatial cues in auscultation. The system combines a tailored residual encoder with a 16-channel–adapted ResNet and aggregates segment embeddings through gated attention to yield patient-level predictions. Training employs task-matched losses and light, label-preserving augmentations. We evaluate the framework on a newly introduced 16-channel dataset (SPSC-HCM-16C, 183 subjects) under three taxonomies: binary (healthy vs. unhealthy), 3-class (healthy/chronic/non-chronic), and 5-class. Using subject-independent (patient-wise) evaluation with a 60/40 train–test split, the proposed method consistently outperforms all baselines, achieving an F1 score of 97.3% with a sensitivity of 99.1% in the binary task, an F1 score of 77.3% in the 3-class task, and an F1 score of 50.6% in the challenging 5-class setting. Transfer learning on the ResNet branch provides up to a 17.6 percentage-point improvement in macro-F1, while simple augmentations further enhance the minority-class recall. The largest gains arise from explicitly modeling multi-site spatial information via the dual-encoder with gated attention, which stabilizes patient-level decisions and improves the detection of rare classes. Task-appropriate objectives further strengthen performance, with second-order polynomial loss proving most effective for the binary and 5-class tasks, and focal loss favoring the 3-class setting. Overall, the results demonstrate that leveraging spatial cues within a patient-level MIL framework enables robust, interpretable, and clinically meaningful respiratory disease classification.