An interpretable machine learning model for early risk stratification of medium-to-giant coronary artery aneurysm in...

Medium-to-giant coronary artery aneurysm (MGCAA) represents the most severe complication of Kawasaki disease (KD) and remains difficult to identify early. Existing risk scores are not tailored for MGCAA and lack external validation. Interpretable machine learning (ML) approaches may improve early risk stratification.

We retrospectively analyzed 443 patients from Fuzhou (development cohort) and 2,334 from Suzhou (external validation). Multiple ML algorithms were compared, and a random forest (RF ranger) model was selected. Six routinely collected predictors were retained, and model interpretability was assessed using SHapley Additive exPlanations (SHAP). Discrimination, calibration, and decision curve analysis (DCA) were conducted. Intercept-only recalibration was applied for external adaptation.

The final model included hemoglobin, time to diagnosis, oral mucosal changes, rash, triglycerides, and neutrophil percentage. The AUC was 0.70 in internal validation and 0.75 in external validation. Recalibration improved calibration-in-the-large. SHAP analysis supported the clinical interpretability of the selected predictors, and DCA indicated potential net benefit within a limited range of low-to-moderate threshold probabilities. The model is available as a web-based tool to support individualized risk estimation.

This interpretable machine learning model enables early risk stratification for MGCAA using six routinely collected clinical variables, and may assist individualized clinical risk assessment in children with Kawasaki disease.

Medium-to-giant coronary artery aneurysms (MGCAA) are the most severe cardiovascular complication of Kawasaki disease and are difficult to predict early, before imaging confirmation. This study proposes an interpretable machine-learning model for MGCAA risk assessment based on six routinely available clinical variables. The model incorporates SHAP-based transparency, external validation, and intercept-only recalibration to improve generalizability across populations. Deployed as a web-based tool, it facilitates individualized risk stratification and may support clinical decision-making related to monitoring intensity and follow-up strategies.