A machine learning-based model to predict intravenous immunoglobulin resistance in Kawasaki disease.

Accurate prediction of intravenous immunoglobulin (IVIG) resistance is crucial for the effective treatment of Kawasaki disease(KD). This study aimed to develop a predictive model for IVIG resistance in patients with Kawasaki disease and to identify the key predictors. The training set underwent cross-validation, and models were constructed using six machine learning algorithms. Model performance was validated through cross-validation, test set evaluation, and two external validation sets evaluation. The model constructed using the random forest algorithm demonstrated the best overall performance among six models. The areas under the receiver operating characteristic curve (AUCs) for 5-fold cross-validation, internal validation, and external validations from Shaoxing and Quzhou were 0.711, 0.751, 0.827, and 0.735, respectively. According to the Shapley additive explanation (SHAP) method, C-reactive protein-to-albumin ratio, prognostic nutritional index, and sex were identified as the most important predictors. Our model demonstrates strong predictive capability for assessing IVIG resistance in Kawasaki disease patients.