Sleep duration, sleep quality, and daytime napping in relation to incident cardiometabolic multimorbidity across three national aging cohorts: a prospective, interpretable machine learning study
Background Sleep disturbances are modifiable factors linked to metabolic dysregulation, adiposity, blood pressure abnormalities, and cardiometabolic risk. Evidence remains limited on whether nighttime sleep duration, sleep quality, and daytime napping predict cardiometabolic multimorbidity (CMM), defined as two or more of hypertension, type 2 diabetes mellitus, coronary heart disease, and stroke, across populations. Methods Using CHARLS, ELSA, and HRS, we examined associations between sleep patterns and incident CMM in prospective aging cohorts. Kaplan–Meier curves and Cox proportional hazards models were used for time-to-event analyses, adjusting for sociodemographic characteristics, physical activity, and clinical covariates. Restricted cubic splines explored nonlinear associations. Mediation analyses assessed whether blood pressure, lipid markers, and C-reactive protein mediated the association between nighttime sleep duration and CMM. Interpretable machine learning evaluated the contribution of sleep variables within the cardiometabolic risk profile. Because sleep measures differed across cohorts, harmonization was conducted at the construct rather than raw-score level, and cross-cohort effect-size comparisons should be interpreted cautiously. Results The analysis included 13,859 participants: 4,145 from CHARLS, 4,268 from ELSA, and 5,446 from HRS. In fully adjusted Cox models, good sleep quality was associated with lower CMM risk than poor sleep quality in CHARLS and ELSA. A similar trend was observed in HRS, although significance appeared only in complete-case sensitivity analysis. Kaplan–Meier curves showed poorer CMM-free survival among participants with poor sleep quality. For nighttime sleep duration, restricted cubic splines suggested a U-shaped association in ELSA before adjustment, but this weakened after full adjustment, and findings were inconsistent across cohorts. In exploratory mediation analyses, some small indirect effects were observed in unadjusted models, but none remained evident after covariate adjustment. In prediction analyses, logistic regression and LightGBM showed stable performance. SHAP identified systolic blood pressure, waist circumference, and BMI as leading predictors, while nighttime sleep duration added predictive information. Conclusion Better sleep quality was associated with lower incident CMM risk in CHARLS and ELSA, with a similar but less robust pattern in HRS. Nighttime sleep duration showed less consistent associations. Machine learning analyses were exploratory and supported risk stratification and feature interpretation rather than clinical screening.