Explainable ML for ACS culprit plaques: a multidimensional CCTA model highlighting hemodynamic increment.

To develop an interpretable, multi-parameter machine learning (ML) model that integrates plaque morphology, composition, perivascular inflammation, and comprehensive hemodynamic descriptors to identify future acute coronary syndrome (ACS) culprit plaques and evaluate the incremental predictive value of hemodynamic parameters. A total of 217 lesions from 88 patients with pre-ACS coronary computed tomography angiography (CCTA) were analyzed. Anatomical features, plaque composition, computed tomography fractional flow reserve (CT-FFR), adverse plaque characteristics, and the fat attenuation index (FAI) were extracted. Hemodynamic metrics derived from computational fluid dynamics (CFD) were also obtained, including wall shear stress (WSS), time-averaged wall shear stress, oscillatory shear index and relative residence time. Feature selection was performed using the least absolute shrinkage and selection operator and Boruta. Four classifiers were trained with cross-validated hyperparameter tuning, and model interpretability was assessed using shapley additive explanations (SHAP). Three progressive models were constructed to quantify the incremental contribution of hemodynamics. Random forest demonstrated the best performance, and the SHAP analysis revealed WSS and FAI as the most influential predictors. The hemodynamic-integrated model (Model 3) showed significantly improved discrimination and reclassification over the FAI-integrated model (Model 2) (Delta AUC = 0.162, P = 0.035; C-index = 0.917; NRI 0.751, P<0.001; IDI = 0.143, P<0.001). This feasibility study suggests that an explainable, multi-parameter ML framework holds promise for identifying ACS culprit plaques, and highlights the potential critical incremental value of comprehensive hemodynamic assessment.