Development and external validation of a non-laboratory model to predict early-onset cardiovascular disease risk: a multicohort machine learning study.

BACKGROUND AND AIMS: This study aimed to develop and externally validate a questionnaire-based model for two related tasks in adults younger than 65 years: classification of incident early-onset cardiovascular disease (EOCVD) among participants without cardiovascular disease at baseline, and time-to-event prediction of all-cause mortality among participants with EOCVD at baseline.

METHODS

Longitudinal data were obtained from CHARLS, HRS, KLoSA, ELSA, and SHARE. The incidence classification analysis included 37,082 participants, and the prognostic survival analysis included 19,478. Eighteen questionnaire or routinely obtainable predictors were used. No laboratory biomarker or imaging measurement was entered into either model, although reported diagnoses may have depended on earlier clinical testing.

RESULTS

CatBoost achieved an AUROC of 0.6869, an AUPRC of 0.1438, a precision of 0.1105, and a recall of 0.6512 in the independent SHARE cohort. The Cox proportional hazards model achieved a test-set C-index of 0.7175; its 3-, 5-, and 7-year AUROCs were 0.7314, 0.7348, and 0.7350.

CONCLUSIONS

P-EOCVD provides questionnaire-based risk stratification for incident EOCVD and subsequent mortality. Its low precision precludes diagnostic use. The model may support first-step screening where laboratory testing is not immediately available, but prospective validation and recalibration are required before use in under-resourced populations.