He Yu, Guo Wan-Liang, Zhang Ming-Chang
Department of Radiology, Children's Hospital of Soochow University, Suzhou, 215025, China.
BMC Pediatr. 2025 Aug 1;25(1):584. doi: 10.1186/s12887-025-05936-7.
Renal damage in closed spinal dysraphism (CSD), primarily linked to neurogenic bladder dysfunction, significantly impacts long-term patient outcomes by increasing the risk of chronic kidney disease. Identifying patients at highest risk for renal damage is essential for implementing early interventions, improving bladder management strategies, and preserving renal function. This study aims to develop an effective machine learning model to predict renal damage in children with CSD.
This retrospective study included 110 children with CSD. We developed four machine learning models (logistic regression, support vector machine, decision tree, and extreme gradient boosting [XGBoost]), and compared their predictive performances. The area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis were used to evaluate predictive performance. The Shapley additive explanations (SHAP) algorithm and Local Interpretable Model-Agnostic Explanations (LIME) were used to interpret the optimal model.
The XGBoost model showed the best predictive performance (AUC = 0.957) among the four machine learning models. Through the SHAP analysis, abnormal radiological lower urinary tract findings, female sex, and high-grade vesicoureteral reflux were identified as the three most influential features in predicting renal damage.
Our study effectively developed a model that accurately predicted renal damage in children with CSD based on the XGBoost algorithm, demonstrating its potential to achieve good predictive performance.
闭合性脊柱裂（CSD）中的肾损伤主要与神经源性膀胱功能障碍有关，通过增加慢性肾病的风险，对患者的长期预后产生重大影响。识别肾损伤风险最高的患者对于实施早期干预、改善膀胱管理策略和保护肾功能至关重要。本研究旨在开发一种有效的机器学习模型，以预测CSD患儿的肾损伤。
这项回顾性研究纳入了110例CSD患儿。我们开发了四种机器学习模型（逻辑回归、支持向量机、决策树和极端梯度提升[XGBoost]），并比较了它们的预测性能。采用受试者工作特征曲线（AUC）下面积、校准曲线和决策曲线分析来评估预测性能。使用Shapley加法解释（SHAP）算法和局部可解释模型无关解释（LIME）来解释最优模型。
在四种机器学习模型中，XGBoost模型表现出最佳的预测性能（AUC = 0.957）。通过SHAP分析，放射学下尿路异常表现、女性性别和高级别膀胱输尿管反流被确定为预测肾损伤的三个最具影响力的特征。
我们的研究基于XGBoost算法有效地开发了一个准确预测CSD患儿肾损伤的模型，证明了其实现良好预测性能的潜力。
