Development and validation of a risk prediction algorithm for recurrent cardiovascular events in the Chinese population: P-CARDIAC.

Abstract

Existing cardiovascular disease (CVD) risk prediction tools may not be applicable to the Chinese populations because of their development based on the mostly Western cohorts and limited list of covariates. To develop and validate a machine learning-based risk prediction model that can be used by primary care physicians in Hong Kong, Personalized CARdiovascular DIsease risk Assessment for Chinese (P-CARDIAC), to predict 10-year CVD risk among the Chinese population. A novel algorithm based on the Hong Kong West Cluster cohort was designed by Cox proportional hazards (CPH) model with LASSO for shortlisting statistically significant risk factors and XGBoost to achieve better performance and interpretability by medical professionals. The internal validation was performed by 100 repeats of 10-fold cross-validation while the external validation was evaluated in two other independent cohorts with the comparison of TIMI risk score and SMART2. A total of 48 799 participants with prior CVD events were included and externally validated by two other Hong Kong cohorts with 119 672 and 140 533 participants. The novel algorithm had better performance than the CPH model with a 0.69 C-statistic. The external validation showed great model calibration and clinical utility with 0.62 and 0.64 C-statistic, respectively, for the two cohorts; while both TIMI and SMART2 were underperforming. Medication treatments also had a strong correlation with recurrent CVD. P-CARDIAC allows a more personalised approach for recurrent CVD prevention with dynamic baseline risk and concurrent medication effect. Such an approach with the potential for being recalibrated for other ethnicities will be used in primary care for managing CVD risk.