Development and validation of an interpretable machine learning model for predicting medium-to-giant coronary...

This study aimed to develop and validate an interpretable machine learning (ML) model using routinely collected clinical data to predict medium-to-giant coronary artery aneurysms (MGCAA) early in Kawasaki disease (KD).

This retrospective study included 2,777 KD patients from two centers in China. Eleven ML algorithms were developed using clinical and laboratory data from electronic medical records (EMRs). Recursive feature elimination and SHapley Additive exPlanations (SHAP) were used for feature selection and interpretability. The final model was internally and externally validated, with intercept-only recalibration to correct miscalibration, and evaluated by area under the receiver operating characteristic curve (AUC), calibration, and decision curve analysis (DCA). The model was deployed as an R Shiny-based online prediction tool.

The support vector machine model implemented with kernlab (SVM kernlab) included seven key features: time to diagnosis, monocyte percentage, rash, eosinophil percentage, C-reactive protein, triglycerides, and neutrophil percentage. It achieved an AUC of 0.732 (95% CI, 0.597-0.866) in internal validation and 0.689 (95% CI, 0.611-0.767) in external validation. SHAP analysis provided both global feature importance and individualized explanations. Recalibration improved model calibration, and DCA demonstrated meaningful net clinical benefit across clinically relevant threshold probabilities.

This study presents an interpretable ML model to predict MGCAA risk in KD using routine clinical data, supporting clinical risk assessment.