Related Experiment Videos
Study of a Kawasaki disease diagnostic prediction model based on the LightGBM machine learning algorithm
Hongyan Li1, Yushan Li2, Chuxiong Gong1
1Department of Cardiology, Kunming Children's Hospital, Kunming, Yunnan, China.
Frontiers in Artificial Intelligence
|June 10, 2026
Summary
Machine learning accurately diagnoses Kawasaki disease (KD) using key biomarkers, improving early detection and reducing coronary artery lesions (CAL). A simplified model with six variables offers high efficacy and clinical applicability for rapid diagnosis.
Area of Science:
- Pediatric rheumatology
- Computational medicine
- Biomarker discovery
Background:
- Kawasaki disease (KD) is a critical pediatric vasculitis with unclear pathogenesis.
- Lack of specific biomarkers complicates early diagnosis, risking coronary artery lesions (CAL).
- Diagnostic uncertainty delays treatment and increases CAL risk.
Purpose of the Study:
- Develop a machine learning (ML) diagnostic model for KD.
- Optimize the KD diagnostic workflow for early identification.
- Reduce the incidence of CAL through timely intervention.
Main Methods:
- Retrospective analysis of 4,469 pediatric patient records (2017-2023).
- Utilized Light Gradient Boosting Machine (LightGBM) for model development.
- Employed SHapley Additive exPlanations (SHAP) for feature importance and model interpretation.
Main Results:
- Full LightGBM model achieved an AUC of 0.9956, accuracy 0.9653, sensitivity 0.9596, specificity 0.9717.
- SHAP identified CRP, APTT, Ca2+, ESR, Cl-, and other variables as key predictors.
- Simplified model using WBC, PLT, CRP, APTT, TT, ALB achieved AUC 0.9792, accuracy 0.9340, sensitivity 0.9277, specificity 0.9410.
Conclusions:
- ML simplifies KD diagnosis, offering high efficacy.
- The simplified model demonstrates robust discriminative capacity and clinical applicability.
- An offline application supports rapid, accurate early KD diagnosis, especially in primary care.