Paediatric respiratory sound classification by integrating local feature extraction and global context modeling

BACKGROUND AND OBJECTIVE: Automated respiratory sound classification based on deep learning is critical for enhancing diagnostic precision and efficiency in respiratory disease. Research on respiratory sound classification for adults is extensive, whereas work on paediatric respiratory sounds is still scarce. Compared to adults, paediatric respiratory sounds typically contain more high-frequency components, and this inherent difference presents unique challenges for models. APPROACH: In response to this challenge, we proposed a novel architecture, which integrated the local feature extraction module with the global context model. Specifically, we designed the Mel_Grouper module to serve as a front-end with the purpose of enhancing local pathological representations. The output was fed into the Transformer-based Mel_Encoder to fuse global context. MAIN RESULTS: Experimental results on the SJTU Paediatric Respiratory Sound (SPRSound) dataset demonstrate that our method achieves state-of-the-art (sota) performance. Regarding the four subtasks of the SPRSound dataset, it outperforms the previous best results by 3.37%, 3.06%, 6.83% and 5.36%, respectively. To better reflect the real clinical conditions, we further performed experiments on the real-world paediatric respiratory sound dataset. SIGNIFICANCE: These results demonstrate the method's significant performance in paediatric respiratory sound classification. The code is available at: https://github.com/shufei2580/SSAST/tree/master.