Explainable deep learning algorithm for distinguishing IVIG-Resistant Kawasaki disease in Shandong peninsula, China

Intravenous immunoglobulin (IVIG) resistance of Kawasaki disease (KD) patients have a heightened risk of coronary artery lesions. We aimed to explore the predictive factors of IVIG resistance of KD from Shandong Peninsula in China, and established a explainable prediction model based on deep learning.

This retrospective cohort study include the patients and were divided into two subgroups: IVIG sensitive and IVIG resistance, then arbitrarily partitioned into training set and validation set at a ratio of 8:2. The data was then trained through support vector machines, Random forest, Decision tree, XGBoost, and LightGBM to determine the optimal model. Split and Gain were used to analyze the importance of variables.

This study included 914 KD patients (768 IVIG-sensitive and 146 IVIG-resistant). Comparative analysis revealed 19 laboratory indicators with significant differences by violin plots. The LightGBM displayed the highest comprehensive predictive capability trained with the above data, with area under curve (AUC) of 0.9936. Applying importance analysis (split and gain), 6 indicators were chosen, including C-reactive protein, Serum sodium, Albumin, Hemoglobin, Neutrophil percentage and Platelet. The LightGBM-Clinic model trained by the above indicators, with AUC of 0.9725.

LightGBM can predict IVIG-resistance well in KD patients in Shandong Peninsula, China. C-reactive protein, Serum sodium, Albumin, HB, Neutrophil percentage, and PLT are important indicators, enhanced explainable of the model.