Study of a Kawasaki disease diagnostic prediction model based on the LightGBM machine learning algorithm

Kawasaki disease (KD) is an acute systemic vasculitis predominantly affecting children under 5 years of age. Its pathogenesis remains incompletely understood, and the lack of specific diagnostic biomarkers during the acute phase poses substantial challenges to clinical diagnosis. Such diagnostic uncertainty often results in missed or delayed cases, leading to lost therapeutic opportunities and the subsequent development of coronary artery lesions (CAL). The present study aimed to establish a machine learning-based diagnostic model to optimize the KD diagnostic workflow, enable early identification, and reduce the incidence of CAL.

We retrospectively analyzed Pre-treatment clinical data from 4,469 patients admitted to Kunming Children's Hospital between January 2017 and December 2023, including 2,345 patients diagnosed with Kawasaki disease (case group) and 2,124 febrile non-KD patients (control group). A Light Gradient Boosting Machine (LightGBM) algorithm was employed to develop a diagnostic model for KD. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), as well as accuracy, sensitivity, and specificity calculated from the confusion matrix at a classification threshold of 0.5. SHapley Additive exPlanations (SHAP) analysis was performed to identify key predictive features and interpret the model's decision-making mechanism. On the basis of SHAP-derived feature importance and clinical availability, six core variables were selected to establish a simplified diagnostic model, which was further deployed as an offline application.

The full-feature LightGBM model achieved an outstanding area under the receiver operating characteristic curve (AUC) of 0.9956. At a classification threshold of 0.5, the confusion matrix yielded an accuracy of 0.9653, a sensitivity of 0.9596, and a specificity of 0.9717. SHAP analysis revealed that C-reactive protein (CRP), activated partial thromboplastin time (APTT), total calcium (Ca²⁺), erythrocyte sedimentation rate (ESR), serum chloride (Cl^-), and several other variables exhibited strong predictive value. To achieve an optimal balance between predictive performance and clinical applicability, six core variables were selected: white blood cell count (WBC), platelet count (PLT), CRP, APTT, thrombin time (TT), and albumin (ALB). The simplified LightGBM model constructed using these parameters achieved an AUC of 0.9792; at the 0.5 classification threshold, it demonstrated an accuracy of 0.9340, a sensitivity of 0.9277, and a specificity of 0.9410, indicating robust discriminative capacity. Finally, an offline executable application was developed to support clinicians-especially those in primary care settings-in achieving rapid and accurate early diagnosis of KD.

Machine learning approaches enable effective simplification of the diagnostic process for KD. The predictive model constructed on the basis of WBC, PLT, CRP, APTT, TT, and ALB shows high diagnostic efficacy and favorable clinical applicability.