Machine learning-based advanced coronary artery disease pretest probability model: Comparison with conventional pretest probability models

Background: Pretest probability (PTP) models using clinical risk factors guide decision-making for coronary artery disease (CAD). Existing models (Updated Diamond-Forrester [UDF] and CAD Consortium [CAD2]) exhibit suboptimal predictive efficacy in Asian populations due to ethnic differences in atherosclerosis and risk profiles. We developed an advanced CAD-specific PTP model using ridge-penalized logistic regression and validated its reliability. Methods: Utilizing data from 4,696 Korean patients (3 trials and 2 cohorts), we employed ridge regression to develop an advanced PTP model (K-CAD) for identifying patients with CAD with >=50% diameter stenosis, determined using coronary computed tomography or invasive coronary angiography. External validation used datasets from another tertiary center (External Validation Cohort 1, n=428) and a nationwide health checkup cohort (External Validation Cohort 2, n=117,294). We compared K-CAD with existing models using continuous receiver operating characteristic (ROC) and ternary net reclassification improvement (NRI) analyses. Findings: Continuous ROC analysis in External Validation Cohort 1 revealed areas under the curves (AUCs) for UDF, 0.68 (95% confidence interval [CI] 0.63-0.73); CAD2, 0.71 (95%CI 0.67-0.76), and K-CAD, 0.76 (95%CI 0.71-0.80). K-CAD significantly outperformed UDF (p <0.001) and CAD2 (p <0.05). NRI analysis demonstrated that K-CAD improved reclassification of non-obstructive patients into low-risk categories. External validation using the nationwide dataset (surrogate endpoint: ICD-10 I20) yielded AUCs for UDF, 0.61 (95% CI 0.58-0.64); CAD2, 0.66 (95%CI 0.63-0.69); and K-CAD, 0.67 (95%CI 0.64-0.70). Interpretation: The study demonstrated K-CAD's utility employing extensive high-quality datasets, highlighting its potential for predicting CAD risk in the Korean population.