SmartOSA-LLM: A large language model enhanced obstructive sleep apnea classification and reporting framework

Obstructive sleep apnea (OSA), as a common public health burden globally, has an independent correlation with increased risk of cardiovascular issues, which has been well recognized. However, the diagnosis rate of OSA in clinics remains significantly lower than its real prevalence, constrained by low screening efficiency and diagnostic subjectivity. Automated approaches have previously had limitations in balancing classification accuracy and interpretability, which limit their clinical application. Here, we propose SmartOSA-LLM, a two-stage framework for OSA severity classification. Firstly, a fine-tuned Large Language Model (LLM) is used to simulate clinician reasoning in identifying complex symptomatic features from heterogeneous data. Suspected cases of OSA are preliminarily screened out. Subsequently, an adaptive weighted ensemble model is applied to integrate LLM outputs with original clinical features to obtain precise classification. We use the Sleep Heart Health Study (SHHS) dataset to validate its performance. The proposed framework achieves an overall accuracy of 87.3% in a five-class classification task and a screening accuracy of 91.3% in a binary classification task, meeting clinically accepted diagnostic accuracy standards. Moreover, interpretable reports can be generated, enhancing transparency and strengthening clinical trust. It can effectively support physicians in making accurate diagnoses and improve their understanding of the diagnostic reasoning process. SmartOSA-LLM offers a scalable framework that demonstrates the feasibility of integrating LLM-based reasoning with traditional clinical features for OSA severity classification. The retrospective validation on the SHHS cohort suggests its potential for future clinical translation, though external multi-center validation and prospective clinical evaluation are required before real-world deployment.