Development and internal validation of a monocyte-to-albumin ratio-based risk assessment model for coronary artery lesions in Kawasaki disease.

INTRODUCTION: Coronary artery lesions (CAL) are the most severe complication of Kawasaki disease (KD), yet accurate and practical risk assessment tools remain limited. This study aimed to develop and validate a parsimonious logistic regression model for CAL risk assessment using routine clinical and laboratory parameters. METHODS: A total of 324 consecutive KD patients were retrospectively enrolled. Candidate predictors were screened by univariate analysis and a strict consensus-based strategy integrating five feature-selection methods (LASSO, random forest, Boruta, recursive feature elimination, and univariate threshold). Ten machine learning algorithms were compared, and the final model was selected based on performance and interpretability. A nomogram, calibration curve, and an online calculator were developed. RESULTS: Among 324 patients, 64 (19.8%) had baseline CAL. Four variables [fever duration, GGT, monocyte-to-albumin ratio (MAR), and ESR] were consistently selected by all five methods. Logistic regression yielded a Brier score of 0.142, with high specificity (0.846), but limited sensitivity (0.421) on the hold-out test set. No significant effect modification was observed in exploratory interaction analyses assessing the association of MAR on CAL across the evaluated covariates (all adjusted P > 0.05). GBM model offered superior sensitivity (0.684) than other models, suggesting a potential exploratory role in sensitivity-driven screening scenarios. An online calculator was deployed strictly for exploratory research purposes. CONCLUSIONS: A simple four-variable logistic model incorporating fever duration, GGT, ESR, and MAR provides preliminary CAL risk assessment. However, the modest event count and single-center nature mandate multi-center external validation before any clinical application.