Machine learning classification of prevalent cardiovascular disease in middle-aged and older Chinese adults with circadian syndrome from CHARLS

Circadian syndrome (CircS) integrates metabolic abnormalities with sleep and psychological disturbances and is increasingly linked to cardiovascular disease (CVD). However, few studies have applied machine learning (ML) to characterize and identify prevalent CVD specifically in individuals with CircS. This cross-sectional study analyzed 4,091 middle-aged and older adults with CircS from the 2015 China Health and Retirement Longitudinal Study (CHARLS). To eliminate data leakage, missing-data multiple imputation (MICE) was fitted strictly on the partitioned training set (70%, n = 2,863) and applied to the held-out testing set (30%, n = 1,228). Several ML algorithms—generalized linear model (GLM), gradient boosting machine (GBM), random forest (RF), deep learning (DL), and a stacked ensemble—were trained and hyperparameter-tuned with 5-fold cross-validation. Discrimination and calibration were rigorously evaluated using bootstrap 95% confidence intervals (CIs) for AUC, Brier scores, calibration slope, calibration-in-the-large, and decision curve analysis (DCA). The unweighted prevalence of self-reported CVD in the analytic sample was 26.9% (1,101/4,091). In the independent testing set, the stacked ensemble achieved an AUC of 0.733 (95% CI: 0.703–0.764) and a Brier score of 0.176, which was comparable to RF (AUC = 0.730, 95% CI: 0.699–0.761; Brier = 0.179) and GLM (AUC = 0.728, 95% CI: 0.697–0.759; Brier = 0.176). The AUC differences between the ensemble and RF or GLM were numerically small. Quantitative calibration showed satisfactory agreement, with calibration slopes of 1.100 (95% CI: 0.938–1.286) for the ensemble and 0.960 (95% CI: 0.818–1.130) for GLM. Dyslipidemia, hypertension, kidney disease, and age were the leading variables contributing to CVD classification across models. DCA showed positive net benefit across threshold probabilities of 10%–70%. ML algorithms, including both transparent linear models and tree-based ensembles, demonstrated acceptable discriminative capacity and favorable calibration for identifying prevalent self-reported CVD among individuals with CircS. Simpler models such as RF and GLM showed comparable utility to complex ensembles, providing feasible tools for clinical risk stratification. Prospective studies are needed to assess whether these variables can identify individuals who subsequently develop CVD.