Diagnosis of sleep-disordered breathing using few-shot learning

Sleep-Disordered Breathing (SDB) is a common and clinically significant disorder characterized by recurrent airflow limitation and oxygen desaturation during sleep, which can lead to serious cardiovascular and metabolic complications. Accurate and early diagnosis of SDB is crucial for timely clinical intervention and risk stratification, yet diagnostic results are often influenced by substantial variations in physicians' clinical experience and diagnostic skills across different regions, particularly in large-scale screening and real-world medical settings. However, existing diagnostic methods based on traditional machine learning or fine-tuned deep models often suffer from limited labeled data, poor generalization in few-shot scenarios, and insufficient exploitation of medical domain knowledge. To address these challenges, in this paper, we propose a few-shot method that integrates prompt learning with contrastive learning for SDB diagnosis, short for SDB-FL. Specifically, SDB-FL employs a manual prompting strategy based on handcrafted templates, together with a knowledgeable verbalizer that incorporates medical domain knowledge, to activate latent domain knowledge embedded in pre-trained language models, thereby enabling effective task adaptation under data-scarce conditions. Meanwhile, contrastive learning is introduced to enhance the discriminative ability of representations by promoting intra-class compactness and inter-class separability at the semantic level. Experimental results on both English and Chinese datasets demonstrate that our SDB-FL consistently outperforms strong baseline methods across multiple few-shot settings.