CMR-derived layer-specific strain to differentiate acute myocarditis from acute myocardial infarction with explainable machine learning

Le résumé fourni par la source

The clinical presentation of acute myocarditis (AM) often overlaps with that of acute myocardial infarction (AMI), posing a diagnostic challenge. This study aimed to develop and evaluate an interpretable machine-learning strategy integrating cardiac magnetic resonance (CMR)-derived myocardial strain and clinical variables to differentiate AM from AMI. This retrospective study included 114 patients (52 with AM and 62 with AMI). All patients underwent CMR with comprehensive left ventricular strain analysis, including global, regional, and layer-specific strain parameters. Feature selection, model tuning, and model-size selection were performed within the training data of a nested cross-validation framework, with held-out outer folds used exclusively for performance evaluation. Multiple machine-learning algorithms were compared, and the most parsimonious strategy without a significant loss of discrimination was selected. A final fixed model was subsequently specified using the complete dataset for model interpretation and research demonstration. The five-feature logistic regression (LR) strategy was selected as the preferred approach. Based on pooled predictions from the held-out outer folds, it achieved a ROC-AUC of 0.873 (95% CI, 0.851–0.892) and a PR-AUC of 0.827 (95% CI, 0.789–0.865). At the selected classification threshold, sensitivity and specificity were 0.700 and 0.868, respectively. The final fixed LR model included age, end-systolic apical subendocardial circumferential strain (ESACS_endo), apical longitudinal strain (ALS), lactate dehydrogenase (LDH), and creatine kinase-MB (CK-MB). SHAP analysis demonstrated the relative contribution and direction of effect of these predictors at both global and individual levels. A parsimonious and interpretable model integrating cine CMR-derived myocardial strain with routinely available clinical variables showed good cross-validated discrimination between AM and AMI without incorporating LGE-derived information. Layer-specific and regional strain provided complementary diagnostic information beyond conventional clinical variables. This approach may serve as a noninvasive adjunct for differentiating AM from AMI and warrants further validation in independent cohorts.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.