Identifying the risk of Kawasaki disease based solely on routine blood test features through novel construction of machine learning models.

Abstract

Kawasaki disease (KD) is a leading cause of acquired coronary vasculitis in children and remains a critical diagnostic challenge among febrile pediatric patients. To support frontline pediatricians with a more objective diagnostic tool, we developed and implemented KDpredictor, a machine learning-based model for KD risk identification. KDpredictor leverages only the routine blood test features, including complete blood count with differential count, C-reactive protein, and alanine aminotransferase. It also takes the lead in using age-calibrated eosinophil, platelet, and hemoglobin results. Trained using the light gradient boosting machine algorithm on clinical data from 1,927 KD cases and 45,274 febrile controls, KDpredictor achieved strong performance metrics (auROC: 95.7%, auPRC: 72.4%, recall: 0.89) on a reserved test set, outperforming previous models by at least 3% in auROC and 39.3% in auPRC. Additional explainable AI analyses revealed that several top predictive features in KDpredictor are consistent with prior clinical findings. We also evaluated KDpredictor on three independent cohorts collected in East Asia (Taiwan and China) during the COVID-19 period. KDpredictor achieves recall values of 90.9%, 83.7%, and 91.7% on KD samples identified in three independent medical centers, respectively, indicating its applicability across independent clinical settings. In summary, KDpredictor demonstrates robust generalizability in KD risk identification across populations by using only standard blood samples independent of clinical symptoms. KDpredictor is freely available at https://cosbi.ee.ncku.edu.tw/KD_under7/.