Dynamic prediction of obstructive sleep apnea progression: a deep learning approach using time-varying smartwatch data

Background Obstructive sleep apnea (OSA) was prevalent yet underdiagnosed, contributing to significant cardiovascular morbidity. Early identification of individuals progressing from moderate- to high-risk OSA was critical for timely intervention. Objective To develop and validate deep learning-based survival models predicting the transition from moderate- to high-risk OSA using longitudinal real-world smartwatch data. Methods We analyzed 2,463,329 weekly aggregated health records, including heart rate, blood pressure, sleep parameters, and atrial fibrillation, from the mAFA II long-term cohort between November 22, 2022, and December 18, 2025 in China, monitored with smartwatch (Huawei Technologies Co., Ltd.). Measurement algorithms were validated previously[1-3]. Moderate-risk OSA was defined as apnea-hypopnea index (AHI) 15–30, high-risk as AHI ≥30. The primary outcome was progression to high-risk OSA with the state-of-the-art deep learning approach (Dynamic-DeepHit)[4]. Dynamic-DeepHit was compared with Cox proportional hazards, random survival forest, and survival support vector machine. Four-fold cross-validation was performed, with performance assessed by C-index. Model interpretability was evaluated via feature importance, permutation importance, and partial dependence. Results Of the 131,081 individuals identified as being at moderate risk for OSA (90.1% male) with smartwatch, 29,772 (22.7%) progressed to high-risk OSA over a mean follow-up of 96±29 days. Dynamic-DeepHit outperformed traditional models, achieving higher C-indices (e.g., 0.79 for combined heart rate and blood pressure) versus Cox regression (C-index ~0.65–0.68) and random survival forest (~0.66–0.69). Integrating multimodal data improved accuracy; combining heart rate, blood pressure, and questionnaire data yielded the highest performance (C-index 0.85) (Figure 1). Key predictors included minimum heart rate nadir, maximum systolic blood pressure, and maximum average heart rate. Decreased minimum heart rate and increased maximum systolic blood pressure were associated with higher transition hazard, typically within a median follow-up of 28 weeks. Conclusion We first modeled the temporal trajectory of OSA risk progression and predicted transition from moderate- to high-risk status. The model's capacity to process time-dependent multimodal data enabled dynamic, personalized risk stratification, supporting early, targeted home-based interventions.