Clinical prediction of intravenous immunoglobulin-resistant Kawasaki disease based on interpretable Transformer model

TL;DR

A new interpretable AI model accurately predicts which children with Kawasaki disease will not respond to standard IVIG treatment, helping doctors make better treatment decisions.

Contribution

The study introduces an interpretable transformer-based model (TabPFN-V2) that outperforms existing methods in predicting IVIG resistance in Kawasaki disease.

Findings

The TabPFN-V2 model achieved 97% accuracy in predicting IVIG resistance in Kawasaki disease.

Key predictors of IVIG resistance include elevated AST, CRP, and neutrophil count, while higher platelet and albumin levels are protective.

The model provides both global and local interpretability, enabling transparent clinical decision-making.

Abstract

Intravenous immunoglobulin (IVIG) has been established as the first-line therapy for Kawasaki disease (KD). However, approximately 10%–20% of pediatric patients exhibit IVIG resistance. Current machine learning (ML) models demonstrate suboptimal predictive performance in KD treatment response prediction, primarily due to their limited ability to effectively process categorical variables and interpret tabular clinical data. This study aims to develop and interpretable transformer-based clinical prediction model for IVIG resistant KD and validate its clinical utility. This retrospective study analyzed clinical records of KD patients from the Affiliated Hospital of North Sichuan Medical College (Nanchong, China) between January 1, 2014 and December 31, 2024. A cohort of 1,578 pediatric KD cases was systematically divided into training and validation sets. Six machine learning algorithms - Random Forest (RF), AdaBoost, Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and Tabular Prior-data Fitted Network version 2.0 (TabPFN-V2) - were implemented with five-fold cross-validation to optimize model hyperparameters. Model performance was rigorously evaluated using seven metrics: accuracy, precision, recall, F1-score, Matthews correlation coefficient (MCC), area under the receiver operating characteristic (ROC-AUC), and area under the precision-recall curve (PR-AUC). The top-performing model was subsequently subjected to interpretability analysis through Shapley Additive Explanations (SHAP) to elucidate feature contributions. The transformer-based TabPFN-V2 model demonstrated superior predictive performance in KD analysis, achieving an impressive validation set accuracy of 0.97. Comprehensive evaluation metrics confirmed its robust performance: precision 0.98, recall 0.97, F1-score 0.98, MCC 0.95, ROC-AUC 0.99, and PR-AUC 0.99. Global interpretability analysis through kernel SHAP methodology identified the ten most influential predictive features ranked by significance: Coronary artery lesions (CAL), Aspartate aminotransferase (AST), C-reactive protein (CRP), whether it was incomplete KD (KDtype), Neutrophil count (N), Platelet count (PLT), Albumin (ALB), age, White blood cell count (WBC) and Hemoglobin (Hb). Local interpretability analysis revealed distinct correlation patterns with IVIG resistance:AST, CRP, and N demonstrated significant positive correlations, where elevated values corresponded to increased IVIG resistance risk; PLT and ALB showed negative correlations, with higher levels associated with reduced resistance probability. Notably, age and WBC parameters demonstrated threshold effects, where optimal cutoff values enabled re-calibration of single-variable predictive scores. This threshold-dependent relationship suggests potential clinical utility in risk stratification protocols.The TabPFN-V2 model, leveraging an interpretable transformer architecture, demonstrates dual clinical utilities in KD management: (1) accurate prediction of IVIG resistance risk, and (2) data-driven support for personalized therapeutic decision-making. This framework enables probabilistic estimation of treatment resistance likelihood while providing transparent feature contribution analyses essential for developing patient-specific management protocols.

Figures