AI-based coronary artery disease stratification integrating clinical, lipidomic, and transcriptomic data

Background Obstructive coronary artery disease (CAD) remains a leading cause of morbidity and mortality worldwide. Current diagnostic strategies rely heavily on imaging modalities such as coronary CT angiography and invasive coronary angiography, which are costly, resource-intensive and not universally accessible. Moreover, the development of non-imaging, multimodal ML models for the diagnosis of obstructive CAD remains limited. Purpose The aim of this study was to develop a machine learning model for the diagnosis of obstructive coronary artery disease using non-imaging data, including demographics, biochemical markers, lipidomics and genetic variants. Methods A supervised ML framework was developed to predict obstructive CAD, assuming >50% stenosis to at least one major branch. The SMARTool cohort was utilized, which includes multimodal clinical, molecular, lipidomics profiling and genetic data from 190 patients. An automated data quality assessment pipeline was applied. Class imbalance was addressed using majority class downsampling within training folds. Model development and evaluation were performed using stratified k-fold cross-validation. Three tree-based ensemble models were evaluated: Random Forest, XGBoost, and LightGBM. Bayesian hyperparameter optimization was conducted to maximize discrimination. To reflect a clinically oriented decision-making framework, classification thresholds were selected to prioritize sensitivity while preserving acceptable specificity. Model performance was summarized using cross-validated accuracy, sensitivity, specificity, and related metrics. Model interpretability was assessed using SHAP. Results A total of 190 participants were included, of whom 18% had obstructive CAD. Among the evaluated algorithms, XGBoost demonstrated the highest discriminative performance (AUC=0.714). The XGBoost model achieved a sensitivity of 88.3% and a negative predictive value of 97.4%, supporting its ability to safely identify patients at low probability of obstructive CAD. SHAP analysis identified alkaline phosphatase, age, sphingomyelin species (SM 42:4; SM 42:3), aspartate aminotransferase, phosphatidylserine (PS 36:1), LDL cholesterol, BMI and several genetic variants among the most influential predictors. The integration of lipidomic and genetic features contributed substantially to model discrimination beyond conventional clinical variables. Conclusions A multimodal, non-imaging machine learning model integrating demographic, biochemical, lipidomic and genetic data demonstrated good discriminatory ability and high sensitivity with excellent negative predictive value for obstructive CAD. These findings support its potential role as a pre-imaging screening tool to optimize referral to advanced cardiac imaging and reduce unnecessary diagnostic procedures. External validation in independent cohorts is warranted.