A CMR-based multimodal model to guide invasive angiography referral in patients with suspected obstructive coronary artery disease

Background Cardiac magnetic resonance (CMR) enables noninvasive myocardial assessment, but translating CMR findings into objective CAG referral decisions remains challenging. Artificial intelligence (AI) offers the potential to integrate multimodal data for precision risk stratification; however, AI-driven models for this purpose are currently lacking. Objective This study sought to develop and validate a CMR-Based multimodal model to predict the need for CAG in patients with suspected CAD, and to establish a two-tiered clinical algorithm for primary care triage and tertiary referral decision-making. Method This retrospective single-center study included consecutive adults (≥18 years) with suspected CAD referred for CMR between January 2022 and December 2024. Exclusion criteria were age <18 years, incomplete key data (>5% missing), end-stage organ failure, advanced malignancy, or CMR contraindications. Of 1,690 screened patients, 1,296 met eligibility criteria. Patients were stratified based on subsequent CAG referral into a derivation cohort (n=239) for model development and internal validation, and a separate cohort (n=1,057) without CAG for prespecified applicability testing (risk distribution analysis and clinical impact assessment). Within the derivation cohort, three logistic regression models were developed and internally validated using stratified 10-fold cross-validation. The primary outcome was the need for CAG (defined as obstructive CAD requiring revascularization). Model performance was assessed by AUROC, calibration, decision curve analysis, and SHAP interpretability. Statistical analyses were performed using R 4.5.2 and SPSS 27.0. Result The best performing multimodal model (Model 4) integrating CMR and laboratory was selected to predict and stratify patients. Subjects with a higher ML risk score (probability >0.70) had a significantly higher likelihood of requiring revascularization (unadjusted OR: 12.45; 95% CI: 8.23–18.86; P<0.001), and this relationship persisted after multivariable adjustment (adjusted OR: 8.67; 95% CI: 5.34–14.12; P<0.001). LGE segment count, Hyper-enhancement-Myo, HDL-C, and glucose were identified as the most important predictors. Conclusion The ML-based multimodal model integrating CMR, clinical data, and LGE radiomics accurately predicted the need for CAG in patients with suspected obstructive CAD. Radiomic features provided incremental predictive value beyond conventional assessment. ML combined with SHAP offered explicit explanations of individualized risk predictions, and the proposed two-tiered algorithm enables scalable risk stratification across primary and tertiary care settings.