Early identification of Kawasaki disease using an interpretable machine learning model with robust external validation and clinical utility assessment

Le résumé fourni par la source

Kawasaki disease (KD) is the leading cause of acquired cardiac disease in children, in whom delayed diagnosis carries a substantial risk of coronary artery aneurysms. Differentiating KD from other paediatric febrile illnesses remains clinically challenging because no single laboratory test confirms the diagnosis. Existing machine-learning (ML) models for KD identification lack independent external validation and consistent calibration reporting. We aimed to develop and externally validate an interpretable ML model for early KD diagnosis based on routine laboratory parameters. We conducted a retrospective two-centre TRIPOD + AI diagnostic accuracy study (Wuhan Children’s Hospital, internal n = 5,261; Hubei Provincial People’s Hospital, external n = 461). Nine ML algorithms were benchmarked on 36 candidate routine variables; LASSO yielded 27 non-zero predictors. Models underwent 10 × 10 repeated cross-validation; selection used OOF performance (max AUC; min Brier; sensitivity ≥ 85%). The external cohort was blinded for one-time transportability evaluation. Calibration was assessed per the Van Calster hierarchy; clinical utility by decision curve analysis; interpretability by SHAP. XGBoost achieved the highest OOF AUC (0.962, 95% CI 0.961–0.964), internal test AUC 0.970, and lowest internal test Brier score (0.065). Blinded external AUC was 0.893 (95% CI 0.863–0.920). At threshold 0.699, internal sensitivity was 85.7% (specificity 96.1%) and external sensitivity 65.8% (specificity 95.0%). A three-zone Bayesian framework raised operational sensitivity to 72.8%; site-specific Platt recalibration further restored it to 84.8% at the cost of specificity reduction (95.2% to 75.4%). Likelihood ratios were PLR = 13.16 and NLR = 0.360. SHAP identified prealbumin, GGT, CK, neutrophils, hsCRP, and WBC as top-six features (39.6% of total credit), with high cross-cohort stability (Spearman ρ = 0.949). A Platt-calibrated XGBoost model selected exclusively on OOF performance delivered the best diagnostic profile among nine algorithms, with blinded external validation demonstrating transportability alongside meaningful calibration drift recoverable by site-specific recalibration. The SHAP signature highlights prealbumin and γ-glutamyltransferase as diagnostically informative biomarkers absent from current AHA criteria. The model is freely accessible at https://lismith0802.shinyapps.io/kawasaki-predictor-v17-xgb/ .

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.