Research on Early Identification Model of Intravenous Immunoglobulin Resistant Kawasaki Disease Based on Gradient Boosting Decision Tree

Abstract

Background: To construct a predictive model for intravenous immunoglobulin (IVIG) resistant Kawasaki disease (KD) based on the gradient boosting decision tree (GBDT), so as to early identify children with IVIG resistance and actively take additional treatment to prevent adverse events. Methods: The case data of KD children hospitalized in the Pediatric Department of Lanzhou University Second Hospital from October 2015 to July 2020 were collected. All KD patients were divided into IVIG responsive group and IVIG resistant group. GBDT was used to explore the influencing factors of IVIG-resistant KD and to construct a prediction model. Then compared with previous models, the optimal model was selected. Results: In the process of GBDT model construction, 80% of the data were used as the test set, and 20% of the data were used as the validation set. Among them, the verification set was used to adjust the hyperparameters in GDBT learning. The model performed best with a hyperparameter tree depth of 5. The area under the curve of the GBDT model constructed based on the best parameters was 0.87 (95% CI: 0.85–0.90), the sensitivity was 72.62%, the specificity was 89.04%, and the accuracy was 61.65%. The contribution degree of each feature value to the model was total bilirubin, albumin, C-reactive protein, fever time, and Na in order. Conclusion: The GBDT model is more suitable for the prediction of IVIG-resistant KD in this study area.