Song Zixia, Ming Hongjun, Liu Bin, Liu Dong
Department of Pediatrics, Beijing Anzhen Nanchong Hospital, Capital Medical University (Nanchong Central Hospital), Nanchong, China.
Department of Pediatrics, The Affiliated Hospital of Southwest Medical University, Sichuan Clinical Research Center for Birth Defects, Luzhou, China.
Transl Pediatr. 2025 Feb 28;14(2):208-221. doi: 10.21037/tp-24-359. Epub 2025 Feb 25.
Kawasaki disease (KD) can lead to coronary artery aneurysms (CAA) in approximately 1 in 5 untreated children despite intravenous immunoglobulin (IVIG) therapy in the acute phase. The aim of this study is to develop and validate an explainable machine learning (ML)-based prediction model for CAA in KD.
This study retrospectively analyzed the clinical data of children diagnosed with primary KD at Nanchong Central Hospital, Sichuan Province between 2015 and 2023. Six models, including support vector machine (SVM), K-nearest neighbors (KNN), least absolute shrinkage and selection operator (Lasso), extreme gradient boosting (XGBoost), random forest (RF), and multilayer perceptron (MLP), based on ML algorithms were developed. The model with optimal performance was validated and the explainable SHapley Additive exPlanations (SHAP) analysis was used.
A total of 327 children diagnosed with KD were included in the training set and validation set. Receiver operator characteristic curve analysis showed that XGBoost based model exhibited an optimal performance among the six models. Moreover, for a given CAA positive sample, the sum of the SHAP values of all variables of XGBoost represented the individual deviation from the mean predicted from the entire dataset.
The XGBoost algorithm-based explainable model might be used to predict the occurrence of CAA in children with KD.
川崎病（KD）在急性期即便接受静脉注射免疫球蛋白（IVIG）治疗，约五分之一未经治疗的儿童仍会发生冠状动脉瘤（CAA）。本研究旨在开发并验证一种基于可解释机器学习（ML）的KD患儿CAA预测模型。
本研究回顾性分析了2015年至2023年期间在四川省南充市中心医院诊断为原发性KD的儿童的临床资料。基于ML算法开发了6种模型，包括支持向量机（SVM）、K近邻（KNN）、最小绝对收缩和选择算子（Lasso）、极端梯度提升（XGBoost）、随机森林（RF）和多层感知器（MLP）。对性能最优的模型进行验证，并采用可解释的夏普利值附加解释（SHAP）分析。
训练集和验证集共纳入327例诊断为KD的儿童。受试者工作特征曲线分析表明，基于XGBoost的模型在6种模型中表现最优。此外，对于给定的CAA阳性样本，XGBoost所有变量的SHAP值之和表示个体与整个数据集预测均值的偏差。
基于XGBoost算法的可解释模型可能用于预测KD患儿CAA的发生。
