Machine learning model predicts clotting risk during CRRT in ESKD patients: a SHAP-interpretable approach

Ensuring fluent extracorporeal circulation and preventing circuit clotting are important for end-stage kidney disease (ESKD) patients undergoing continuous renal replacement therapy (CRRT). This study aimed to develop a predictive model using machine learning (ML) algorithms to evaluate clotting risk after initiating CRRT, enhancing treatment safety and effectiveness. This study involved 636 ESKD patients who underwent CRRT. Feature selection was conducted <i>via</i> the least absolute shrinkage and selection operator (LASSO) algorithm. ML algorithms, including support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF), gradient boosting machine (GBM), decision tree, and logistic regression (LR), were applied to construct models through tenfold cross-validation. Model performance was assessed <i>via</i> the area under the receiver operating characteristic curve (AUC) and additional metrics. The Shapley additive explanation (SHAP) values quantify each feature’s contribution. This study included 199 patients with blood clots during extracorporeal circulation, corresponding to an incidence rate of 31.3%. The AUC values were 0.864 (SVM), 0.815 (XGBoost), 0.806 (GBM), 0.778 (RF), 0.732 (Decision Tree), and 0.717 (LR). The SVM exhibited the best performance. The initial dose of low-molecular-weight heparin (LMWH) was identified as the most significant factor influencing coagulation. ML serves as a reliable tool for predicting the risk of extracorporeal circuit clotting in ESKD patients undergoing CRRT. The SHAP method elucidates key risk factors, providing a basis for early clinical intervention.

对于接受连续性肾脏替代治疗（continuous renal replacement therapy, CRRT）的终末期肾病（end-stage kidney disease, ESKD）患者而言，维持体外循环流畅、防止管路凝血至关重要。本研究旨在借助机器学习（machine learning, ML）算法构建预测模型，以评估CRRT启动后的凝血风险，进而提升治疗的安全性与有效性。本研究共纳入636名接受CRRT治疗的ESKD患者，通过最小绝对收缩和选择算子（least absolute shrinkage and selection operator, LASSO）算法完成特征选择。本研究采用支持向量机（support vector machine, SVM）、极端梯度提升（extreme gradient boosting, XGBoost）、随机森林（random forest, RF）、梯度提升机（gradient boosting machine, GBM）、决策树及逻辑回归（logistic regression, LR）等多种机器学习算法构建模型，并通过十折交叉验证开展训练。模型性能以受试者工作特征曲线下面积（area under the receiver operating characteristic curve, AUC）及其他评估指标进行评价，同时采用夏普利可加解释（Shapley additive explanation, SHAP）值量化各特征的贡献度。本研究中有199名患者在体外循环期间发生凝血，发生率为31.3%。各模型的AUC值分别为：支持向量机0.864、极端梯度提升0.815、梯度提升机0.806、随机森林0.778、决策树0.732、逻辑回归0.717，其中支持向量机模型表现最优。低分子量肝素（low-molecular-weight heparin, LMWH）的初始给药剂量被确定为影响凝血状态的最关键因素。机器学习可作为可靠工具，用于预测接受CRRT治疗的ESKD患者的体外循环管路凝血风险；SHAP方法可阐明关键风险因素，为临床早期干预提供依据。