目的 开发并验证基于多种机器学习算法的自体动静脉内瘘（autologous arteriovenous fistula，AVF）闭塞风险预测模型，为临床早期识别高危患者并指导护理干预提供精准、高效的决策支持工具。方法 该研究为回顾性分析，采用便利抽样法纳入2021年8月至2024年1月于江苏省某三级甲等医院接受血液透析治疗的403例患者。根据随访期间是否发生AVF闭塞，将患者分为闭塞组（n=97）与非闭塞组（n=306）。首先采用单因素分析和二元Logistic回归筛选AVF闭塞的独立危险因素，随后应用随机森林、支持向量机、AdaBoost等7种机器学习算法构建预测模型。采用10折交叉验证评估并优化模型性能，并使用SHAP（SHapley Additive exPlanations）算法对随机森林模型进行可解释性分析。结果 Logistic回归分析表明，内瘘使用时间、冠心病史和内瘘位置是AVF闭塞的独立危险因素（P<0.05）。在构建的7种机器学习模型中，AdaBoost模型与随机森林模型展现出最优的综合预测效能，其在测试集上的受试者操作特征曲线下面积分别为0.882和0.876。SHAP可解释性分析揭示，内瘘使用时间、血钙水平、糖尿病史及内瘘位置是对模型预测结果影响最大的4个特征。结论 基于AdaBoost和随机森林算法构建的AVF闭塞风险预测模型具有优良的区分度和准确性，能有效预测血液透析患者的AVF闭塞风险，为个体化护理干预提供了可靠依据。

Objective To develop and validate a risk warning model for autologous arteriovenous fistula（AVF） occlusion using multiple machine learning algorithms，aiming to provide a precise and efficient decision-support tool for early clinical identification of high-risk patients and guidance for nursing interventions. Methods A retrospective analysis was conducted using convenience sampling，including 403 patients undergoing hemodialysis at a tertiary Grade-A hospital in Jiangsu Province from August 2021 to January 2024. Patients were divided into an occlusion group（n=97） and a non-occlusion group（n=306） based on the occurrence of AVF occlusion during follow-up. Independent risk factors for AVF occlusion were first identified using univariable and binary logistic regression analyses. Subsequently，seven machine learning algorithms，including Random Forest，Support Vector Machine （SVM），and AdaBoost，were employed to construct prediction models. Model performance was evaluated and optimized using 10-fold cross-validation，and the best-performing model was interpreted using the SHAP（SHapley Additive exPlanations） algorithm. Results Logistic regression analysis revealed that AVF usage time（AVF time），a history of coronary artery disease（CAG），and AVF location were independent risk factors for occlusion（P<0.05）. Among the seven machine learning models，the AdaBoost and Random Forest models demonstrated the best overall predictive performance，with areas under the receiver operating characteristic（ROC） curve（AUC） of 0.882 and 0.876 on the test set，respectively. SHAP interpretability analysis indicated that AVF usage time，serum calcium level，history of diabetes mellitus（DM），and AVF location were the four most influential features for the model’s predictions. Conclusion The AVF occlusion risk warning models developed based on AdaBoost and Random Forest algorithms exhibit excellent discrimination and accuracy. They can effectively predict the risk of fistula occlusion in hemodialysis patients，providing a reliable basis for personalized nursing interventions.