Integrative machine learning and Mendelian randomization identify causal laboratory biomarkers for coronary artery lesions in Kawasaki disease: a prospective study.

Kawasaki disease (KD) patients could develop coronary artery lesions (CALs) which threatens children's life. We aimed to develop and validate an artificial intelligence model that can predict CALs risk in KD patients. A total of 506 KD patients were included at Children's Hospital of Fudan University. Seven predictive features were identified for model building. Among different machine learning (ML) models tested, Multi-Layer Perceptron Classifier (MLPC), Random Forest (RF) and Extra Tree (ET) demonstrated optimal performance. These were finally chosen for time-across validation. Among three of them, MLPC stands out with its highest accuracy. Besides, Mendelian randomization (MR) analysis also provided genetic evidence. Among seven predictive features, two of them were identified as causal associations with CALs. They are activated partial thromboplastin time (APTT) and red cell distribution width (RDW). The causal mechanism reinforced the biological plausibility of the model. ML-based prediction models, combined with genetic validation through MR, offer a reliable approach for early CALs risk stratification in KD patients. This strategy may facilitate timely clinical interventions.