Predicting coronary artery abnormalities in Kawasaki disease: Model development and validation.

Area of Science:

• Pediatric Cardiology
• Rheumatology
• Medical Imaging

Background:

• Coronary artery abnormalities (CAA) are a significant complication of Kawasaki disease.
• Routine echocardiography at one month post-diagnosis is the current standard for screening CAA.
• Identifying low-risk patients who could potentially avoid routine echocardiography is an ongoing clinical need.

Purpose of the Study:

• To develop and validate predictive models for CAA in Kawasaki disease.
• To assess if routine echocardiography screening could be safely reduced in low-risk individuals.
• To compare the efficacy of simple, logistic regression, and machine learning models in predicting CAA.

Main Methods:

• Utilized two prospective Japanese multicenter registries (PEACOCK and Post-RAISE) for development and external validation.
• Included variables collected within one week of diagnosis to predict CAA (defined as Zmax ≥ 2) at one month.
• Developed and evaluated simple, logistic regression, LightGBM, and XGBoost models, assessing discrimination, calibration, and clinical utility.

Main Results:

• Analyzed 4,973 (PEACOCK) and 2,438 (Post-RAISE) patients; CAA incidence was 5.5% and 6.8%, respectively.
• A simple model using the week 1 maximum Z score achieved an AUC of 0.79 in external validation.
• Adding variables or using complex models (LightGBM, XGBoost) did not significantly improve AUC (increase ≤ 0.02), and models could not safely reduce echocardiographic examinations.

Conclusions:

• Current predictive models, including machine learning approaches, do not sufficiently enhance prediction accuracy over simpler methods to reduce routine echocardiography.
• The established practice of routine echocardiography at one month post-diagnosis for Kawasaki disease patients should be maintained.
• Further research is needed to identify superior predictors for CAA to potentially optimize screening strategies.