Predicting cardiovascular outcomes in Chinese patients with type 2 diabetes by combining risk factor trajectories and machine learning algorithm: a cohort study.

Summary

In 16,378 Chinese adults with type 2 diabetes, a machine-learning model using 4-year trajectories of risk factors achieved C-index 0.80 for 10-year cardiovascular events, outperforming China-PAR and PREVENT with marked NRI/IDI gains. Dynamic trajectories and ML both contributed to improved discrimination and stratification.

Key Findings

• ML-CVD-C achieved C-index 0.80 (95% CI 0.78–0.82) for 10-year CV outcomes, outperforming China-PAR, PREVENT, and a baseline-only ML model (C=0.62–0.65).
• Substantial reclassification improvements versus comparators (NRI 44–58%; IDI ~8–10%), with both trajectories and ML contributing to performance.
• All models tended to overestimate high-risk prevalence; transition analyses showed risk reductions for those remaining stable or reclassified downward.

Clinical Implications

Incorporation of longitudinal risk trajectories into EHR-enabled ML tools could refine cardiovascular risk stratification, guide therapy intensity, and monitor risk transitions in type 2 diabetes.

Limitations

• Single-cohort internal validation without external validation; potential overfitting/generalizability concerns.
• Observed overestimation of high-risk prevalence across models; calibration optimization needed before clinical deployment.

Future Directions

External validation across diverse Chinese and non-Chinese T2D populations; EHR integration trials to test clinical utility and impact on outcomes; continuous calibration updating.