Semi-Supervised Federated Learning for Open-Set Respiratory Sound Classification

초록

Deep learning approaches for respiratory sound classification have shown promise in supporting clinical diagnosis, yet their deployment across multiple institutions faces four concurrent challenges: label scarcity due to the high cost of expert annotation, data heterogeneity arising from different recording equipment and patient populations, the emergence of previously unseen respiratory abnormalities at inference time, and regulatory constraints that prohibit sharing patient data across institutions. Existing methods address these challenges individually but fail to tackle them jointly. This paper proposes the Semi-Supervised Open-Set Heterogeneous Federated Learning (SOHFL) framework. SOHFL integrates prototype-based single-round federated communication with Gaussian mixture model-based test-time adaptation to achieve open-set learning in heterogeneous medical environments without exchanging raw data. Comprehensive experiments across four heterogeneous respiratory sound datasets with five random seeds and seven baselines demonstrate the effectiveness of SOHFL under label-scarce conditions where unlabeled samples outnumber labeled ones. Under these conditions, SOHFL achieves accuracy comparable to the best multi-round baseline while requiring 400 times less communication and 54 to 264 times less training time. Notably, SOHFL achieves up to a 9.6 percentage-point improvement on datasets containing high proportions of unknown samples. Test-time adaptation reduces performance variance by a factor of 4.7, providing predictable performance guarantees critical for clinical deployment. The single-round communication design makes the framework practical for bandwidth-constrained hospital networks.

키워드