Dual-Channel CNN-LSTM with MFCC-Chroma Attention Fusion for Lung Sound Disease Classification

Lung auscultation remains highly subjective, with pulmonologists averaging only 36.5% correct detection rates. This study proposes a dual-channel CNN-LSTM architecture with a constrained MFCC-Chroma gated fusion mechanism for multi-class respiratory sound classification. HINGA V1, the first localized Filipino lung sound dataset, was collected from 51 patients across four Baguio City institutions and combined with the ICBHI 2017 benchmark. The model achieved a macro F1 of 0.4056 across six disease classes on unseen data, with the fusion mechanism consistently favoring MFCC-derived temporal features, supporting a viable path toward a clinically relevant auscultation tool for the Philippine healthcare setting.