3D Cardiac Magnetic Resonance Substrate Features-Based Machine Learning Model for Post-myocardial Infarction Risk Stratification: a Multicenter Study.
Authors
Affiliations (9)
Affiliations (9)
- Department of Radiology, The Second Affiliated Hospital of Kunming Medical University, Kunming, China.
- Department of Ultrasound Diagnosis, Daping Hospital, Army Medical University, Chongqing, China.
- Department of Statistics and Actuarial Science, School of Computing and Data Science, The University of Hong Kong, Hong Kong.
- Department of Radiology, The Six Affiliated Hospital of Kunming Medical University, Yuxi, China.
- Department of Radiology, Yan'an Hospital of Kunming City, Kunming, China.
- Department of Radiology, Hospital of Honghe State Affiliated to Kunming Medical University, Kunming, China.
- Department of Radiology, Qujing No.1 Hospital, Qujing, China.
- Department of Radiology, The Six Affiliated Hospital of Kunming Medical University, Yuxi, China. Electronic address: [email protected].
- Department of Radiology, The Second Affiliated Hospital of Kunming Medical University, Kunming, China. Electronic address: [email protected].
Abstract
Accurate risk stratification post-myocardial infarction (MI) remains challenging. This study aimed to develop interpretable machine learning (ML) models integrating 3D cardiac magnetic resonance (CMR) substrate features to predict major adverse cardiovascular events (MACE) after MI. This retrospective study included MI patients who underwent CMR between May 2015 and October 2024. The primary endpoint was MACE. External validation used multi-center datasets. 3D features (core scar, border zone, abnormal corridors) were extracted via ADAS 3D. Feature selection involved univariate logistic regression and Boruta algorithm. Eight ML models were rained; the top performer, TabPFN, was used to build multimodal models. SHAP analysis provided interpretability. 292 MI patients were finally enrolled. During 35-month median follow-up, 91 experienced MACE. Nine key predictors were identified: three clinical (HDL, chronic kidney disease, tricuspid regurgitation), two functional (Left ventricular ejection fraction, Left ventricular circumferential strain), and four 3D substrate features (border zone mass, corridor mass, burden, and length). Model 4 (clinical + 3D features) showed strong performance across training (Area under the curve [AUC] = 0.91), internal (AUC = 0.82), and external (AUC = 0.89) sets. Model 3 (only 3D features) had an external AUC of 0.90, surpassing clinical (AUC = 0.63) and functional (AUC = 0.49) models. Decision curve analysis highlighted the clinical benefit of incorporating 3D features. SHAP analysis identified corridor mass and burden as key predictors. ML models using 3D CMR substrate features significantly improve post-MI MACE prediction compared to traditional methods, offering interpretable and personalized risk stratification tools.