Multi-modal Machine Learning for CCTA Referral Decision: Integrating Opportunistic Coronary Calcium Screening from LDCT with Clinical Data

Lung cancer screening CT presents an unparalleled opportunity for opportunistic cardiovascular risk assessment via coronary artery calcification (CAC), yet determining when asymptomatic individuals warrant coronary CT angiography (CCTA) remains challenging. We developed and validated a multi-modal machine learning framework that integrates CAC metrics automatically derived from screening CT with comprehensive health check data to guide CCTA referral decisions in 476 asymptomatic adults. Using a validated nnU-Net model for CAC segmentation and training nine algorithms plus ensemble methods on thirty-four clinical variables and three imaging metrics, we achieved a best AUC of 0.911 with Random Forest and a top accuracy of 0.896 with a Multi-Layer Perceptron on an independent test cohort. Ablation experiments demonstrated the supremacy of imaging-only ensembles (AUC = 0.966, accuracy = 0.917) for the imaging-derived CAD-RADS ≥ 3 end point, whereas SHAP analysis confirmed CAC scores as the dominant predictors. This framework enables actionable CCTA triage from lung cancer screening CT combined with routine health data, adding significant cardiovascular risk stratification value to existing screening paradigms without additional scans or cost.