A Machine Learning Model Using Cardiac CT and MRI Data Predicts Cardiovascular Events in Obstructive Coronary Artery Disease.

Summary

In 2038 patients with obstructive CAD followed for a median of 7 years, a multimodal ML model integrating CCTA and stress CMR achieved an AUC of 0.86 for MACE, outperforming established clinical scores and single-modality approaches. The model was externally tested on two independent datasets, supporting generalizability.

Key Findings

• Multimodal ML (CCTA + stress CMR) predicted MACE with AUC 0.86.
• Performance exceeded ESC (0.55), QRISK3 (0.60), Framingham (0.50), segment involvement score (0.71), CCTA alone (0.76), and CMR alone (0.83).
• External testing on two independent datasets supported generalizability.
• Cohort included 2038 patients (mean age 70), with 13.8% MACE over median 7 years.

Clinical Implications

Clinicians could adopt multimodal imaging ML risk scores to better stratify obstructive CAD patients for preventive therapies, revascularization planning, and follow-up intensity. Integration into clinical workflows may reduce reliance on less discriminative clinical risk scores.

Limitations

• Retrospective design with potential residual confounding
• Model interpretability and deployment details not fully described; need for prospective impact studies

Future Directions

Prospective multicenter impact trials to test clinical decision support integration and effect on outcomes; calibration across diverse populations; exploration of model interpretability and treatment selection utility.