O1003: Web Server-Based Machine Learning-Driven Predictive Models for IVIG Resistance in Kawasaki Disease: Clinical Applications and Interpretability

Background: Kawasaki disease (KD) is a leading cause of acquired heart disease in children, with intravenous immunoglobulin (IVIG) resistance affecting 10-20% of cases and significantly increasing the risk of coronary complications. Existing predictive models often lack regional specificity, fail to integrate multiple machine learning (ML) algorithms, and are not optimized for real-time clinical application. This study aims to develop a region-specific, interpretable ML-based model for early IVIG resistance prediction in KD.

Methods: A retrospective cohort of 463 children diagnosed with KD at Fuzhou University Affiliated Provincial Hospital (2012–2024) was analyzed. After data preprocessing—including multiple imputation and feature selection—the dataset was split into training (70%) and testing (30%) sets. Thirteen ML algorithms were implemented for IVIG resistance prediction, including Random Forest (RF), Gradient Boosting Machine (GBM), Support Vector Machine (SVM), Generalized Linear Model (GLMNet), K-Nearest Neighbors (KNN), Decision Tree, RF Ranger, Neural Network (NN), Lasso Regression, Naïve Bayes, Logistic Regression, and AdaBoost. These models were evaluated via cross-validation. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value, and negative predictive value. SHapley Additive exPlanations (SHAP) analysis was applied to determine feature importance, and an optimal probability threshold was established for clinical risk stratification.

Results: The RF model demonstrated the highest predictive performance (AUC = 0.78). Based on SHAP analysis, a final explainable model incorporating 10 key features—glucocorticoid use, serum creatine kinase (CK), eosinophil percentage (EOS), glucose (GLU), serum creatine kinase-MB (CKMB), gamma-glutamyl transferase (GGT), serum creatinine (Cr), neutrophil count (Neu), Age, and C-reactive protein (CRP)—was developed. To enhance clinical utility, a web-based tool integrating the Youden index (16.9%) was deployed for real-time IVIG resistance risk estimation.

Conclusion: This study presents a region-specific, interpretable ML model for IVIG resistance prediction in KD. By incorporating a user-friendly web-based application, it enhances predictive accuracy and clinical applicability, providing a practical tool for early risk stratification and personalized treatment. Its implementation may facilitate timely therapeutic decisions and improve patient outcomes.