Explainable Machine Learning for Kawasaki Disease Prediction Using Embedded Feature Selection and SHAP-Based Interpretability Analysis

Kawasaki disease (KD) is an acute systemic vasculitis principally affecting children below the age of 5 years, characterized by persistent fever and possible cardiovascular complications. Due to nonspecific clinical presentations and lack of validated biomarkers, its early diagnosis remains a significant challenge. This study proposes a machine learning framework for KD prediction, incorporating embedded feature selection via gradient boosting machines, multi-algorithm ensemble modeling, and SHAP-based explainability analysis. A dataset comprising 645 patients with 84 clinical and laboratory features was obtained from the Pediatric Heart Network. Six classifiers were evaluated: Logistic Regression, LightGBM, Gradient Boosting, Explainable Boosting, AdaBoost, and Decision Tree. Logistic Regression achieved superior performance with an AUC-ROC of 91%, accuracy of 87%, precision of 88%, recall of 91%, and F1-score of 89%. Key predictive features included cervical lymphadenopathy, edema of the hands, and rash, with SHAP analysis confirming their clinical relevance. The framework integrates feature selection with in-terpretability, equipping clinicians with transparent and trustworthy diagnostic support. This work demonstrates the potential of explainable AI to enhance early KD diagnosis and clinical decision-making.