Wearable-derived REM sleep as a modifiable risk factor for cardiovascular disease: A multimodal prediction model in...

Sleep is increasingly recognized as a modifiable factor for cardiovascular diseases (CVDs), yet the roles of specific sleep stages, particularly rapid eye movement (REM) sleep, and the utility of wearable-derived sleep parameters for CVD prediction remain unknown. Therefore, we aimed to investigate the associations of sleep with CVD risk and to develop multimodal prediction models for incident CVD. Using data from the All of Us Research Program, 23,413 adults who consented to share both Fitbit and electronic health record data between May 2018 and October 2023 were included in the analysis. Sleep data were obtained from Fitbit wearable devices, including sleep duration and stage-specific information such as REM sleep. Incident CVD was defined as occurring at least six months after the initiation of Fitbit monitoring. Cox proportional hazards models were applied to estimate adjusted hazard ratios (aHR) with 95% confidence intervals (CIs) for sleep duration and REM sleep in association with CVD. For the prediction of CVD within six years, two modeling strategies were implemented, including a non-invasive multimodal model using demographic, self-reported, and wearable-derived data and an invasive multimodal model that additionally incorporated laboratory measures. To compare real-world model performance, we applied and compared several conventional risk prediction models, including SCORE2, the Framingham Risk Score, AutoPrognosis 2.0, and the Pooled Cohort Equations. Among 23,413 participants (mean [standard deviation] age, 56.7 [15.5] years; 70.9% female), both short (<7 h) and long (≥9 h) sleep durations were associated with higher risk of CVDs compared with 7 to 9 h of sleep (short sleep: aHR, 1.16 [95% CI, 1.07 to 1.27]; long sleep: aHR, 1.32 [95% CI, 1.10 to 1.58]), with U-shaped relationship. Individuals with less than 20% REM sleep had a greater risk of CVDs (aHR, 1.31 [95% CI, 1.18 to 1.45]) compared with those with 20% to 25% REM sleep. These associations were consistent across CVD subtypes. In prediction analyses, soft-voting ensemble models with LightGBM and XGBoost integrating wearable, clinical, and laboratory data achieved high performance (AUROC of non-invasive models, 0.748; AUROC of invasive models, 0.782), outperforming conventional risk scores (SCORE2, the Framingham Risk Score, AutoPrognosis 2.0, and the Pooled Cohort Equations: AUROC range, 0.649 to 0.685). Short and long sleep durations, as well as reduced REM sleep, were associated with increased risk of CVD. We also derived and validated a multimodal model to predict the incidence of CVD within six years. These findings support the potential value of integrating objective sleep-stage measures and multimodal digital health data into future cardiovascular prevention strategies and guideline development.