CardioHolisticAI for multimodal deep learning based coronary artery disease risk stratification and personalized treatment recommendation

Despite the significant reduction in mortality and morbidity caused by coronary artery disease (CAD), early stratification of risk is important for timely intervention and individualised care. Most current CAD prediction methods rely on clinical variables, focus on single-modality analysis or simple fusion, and lack interpretability, calibration, and robustness assessment. The authors introduce CardioHolisticAI, a multimodal deep learning framework that also features MultiRiskNet for CAD risk stratification in Low, Medium, and High risk. MultiRiskNet leverages modality-specific encoders and attention-guided fusion to combine clinical features, modality-agnostic radiomic descriptors, and genomic pathway information while maintaining modality-specific representations. A multilevel explainability module integrates SHAP-based global importance, Integrated Gradients-based local attribution, radiomic feature ranking, and pathway-level genomic analysis. MultiRiskNet was evaluated with repeated five-fold cross-validation and achieved an accuracy of 98.50 ± 0.16%, a Macro-F1 score of 0.985 ± 0.0015, and a one-vs-rest ROC-AUC of 0.992 ± 0.0009, outperforming all evaluated unimodal and conventional fusion baselines. Noise injection, feature masking, and repeated-initialisation experiments suggest stable performance under the assessed internal conditions, and calibration analysis examined the reliability of the predicted probabilities. This is because the imaging source is non-cardiac and the source modalities are independent and not clinically validated at the patient level. Before clinical deployment, external evaluation in a patient-linked multimodal cardiac cohort is needed. Not applicable.