Machine learning model for intravenous immunoglobulin resistance in Kawasaki disease: model development and validation study

Background Intravenous immunoglobulin (IVIG) resistance in Kawasaki disease (KD) increases coronary artery risk. Early prediction is crucial for improving outcomes. This study aimed to develop and validate a machine learning (ML) model for predicting IVIG resistance in children with KD. Methods A retrospective cohort of patients with KD was used for model development, with external validation cohorts from Fuzhou and Yangzhou, and a prospective validation cohort. Clinical and laboratory variables were extracted from electronic medical records. We evaluated 12 algorithms and support vector machine (SVM) was selected for optimal performance. SHapley Additive exPlanations (SHAP) values assessed feature importance, followed by stepwise feature elimination. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). A web-based calculator was developed. Results A total of 2371 patients with KD involved in the retrospective development cohort, 443 in Fuzhou cohort, 198 in Yangzhou cohort, and 253 in prospective validation cohort. The SVM model achieved AUCs of 0.782 in internal validation, 0.746 and 0.759 in the external validation cohorts from Fuzhou and Yangzhou, respectively, and 0.799 in prospective validation. The final model incorporated eight predictors, with SHAP analysis providing both global and local explanations of feature contributions. The model also demonstrated good calibration and favorable net benefit across clinically relevant thresholds in DCA. Conclusions The SVM-based ML model using routine clinical data shows potential for predicting IVIG resistance in KD and may support early risk stratification.