DAM-AST: An adaptive feature fusion and multi-scale attention enhancement network for respiratory sound-based respiratory disease classification

Abstract Respiratory diseases are one of the major global health challenges, and early detection and timely intervention are crucial for improving patient outcomes. In recent years, automated classification methods for breath sounds have been extensively studied, offering novel approaches for assisting the diagnosis of respiratory disorders. However, current methods have evolved beyond single-feature extraction, continuously optimizing multi-feature fusion and detail-capture capabilities. To address the limitations of traditional respiratory sound classification methods in these aspects, this paper proposes a Delta-feature-driven Adaptive Fusion and Multi-Scale Attention Network (DAM-AST). The study extracts static Log-Mel spectral features and dynamic Delta difference features from respiratory sounds while designing a dual-channel synchronous random masking strategy to enhance robustness against noise and signal loss. An Adaptive Feature Fusion Module (AFFM) automatically learns the weighted ratio of static and dynamic information to achieve deep feature complementarity, and a Multi-Scale Hierarchical Attention Enhancement (MSHA) Module applies channel and spatial attention at different scales to balance global trends with local details. A pre-trained AST model serves as the backbone network and is fine-tuned for classification tasks. The proposed method achieves excellent classification performance on fused data across two datasets, and t-SNE visualization and Attention Rollout analysis further demonstrate that the model forms distinct feature clusters across different breath sound categories, providing a robust foundation for the classification task.