中国血液净化

中国血液净化 ›› 2026, Vol. 25 ›› Issue (04): 278-283.doi: 10.3969/j.issn.1671-4091.2026.04.002

• 临床研究 • 上一篇    下一篇

基于不同算法的预测模型在维持性血液透析短期内死亡风险预测中的比较分析

张苗苗   吕红红   

  1. 710038 西安，空军军医大学第二附属医院肾内科
  • 收稿日期:2025-08-28 修回日期:2025-12-24 出版日期:2026-04-12 发布日期:2026-04-12
  • 通讯作者: 吕红红 E-mail:drsr008@163.com
  • 基金资助:
    陕西省医学科学研究课题计划(2021JM6324)

Comparative analysis of prediction models based on different algorithms in the prediction of the risk of death in the short term of maintenance hemodialysis for uremia

HANG Miao-miao, LV Hong-hong   

  1. Department of Nephrology, the Second Affiliated Hospital of Air Force Medical University， xi'an 710038, China
  • Received:2025-08-28 Revised:2025-12-24 Online:2026-04-12 Published:2026-04-12
  • Contact: 空军军医大学第二附属医院肾内科 E-mail:drsr008@163.com

摘要： 目的  本研究通过比较不同机器学习算法，筛选维持性血液透析(maintenance hemodialysis for uremia，MHD)患者短期死亡风险的最优预测模型，为临床早期风险分层与干预提供依据。方法  回顾性纳入2022年4月─2024年5月在空军军医大学第二附属医院接受MHD治疗的197例尿毒症患者，根据12个月内生存结局分为死亡组(n=54)和存活组(n=143)。基于多因素Logistic回归分析筛选出的与短期死亡相关的独立影响特征，构建逻辑回归(logistic regression，LR)、随机森林(random forest，RF)、极端梯度提升(extreme gradient boosting，XGBoost)、支持向量机(support vector machine，SVM)和K近邻算法(k-nearest neighbors，KNN)等5种机器学习预测模型。通过受试者工作特征曲线下面积(area under the curve，AUC)、灵敏度、特异度、F1分数及准确率综合评价模型性能。利用决策曲线分析(decision curve analysis，DCA)评估临床净获益，并应用Shapley加性解释(Shapley additive explanations，SHAP)方法量化各临床特征对模型预测的贡献度。 结果  多因素Logistic回归结果显示：糖尿病(OR=14.347、95%CI：1.796～114.624、P=0.012)、总胆固醇(OR=62.811、95%CI：2.463～1601.852、P=0.012)、C反应蛋白(c-reactive protein，CRP)(OR=4.723、95%CI：1.592～14.011、 P=0.005)、 白蛋白(albumin,ALB)(OR=-0.917、95%CI：0.231～0.693、P=0.001)为尿毒症MHD患者短期内死亡的独立影响因素。5种机器学习预测模型性能比较结果显示，KNN模型的AUC值高达0.909，F1分数为0.796，被确定为尿毒症MHD患者短期内死亡风险的最佳预测模型。SHAP解释性分析结果发现，ALB、CRP、糖尿病和总胆固醇依次成为影响模型预测效能的最重要特征。 结论  KNN模型整合ALB、CRP、糖尿病与总胆固醇等关键指标，可有效预测尿毒症MHD患者的短期死亡风险。

关键词: 维持性血液透析, 死亡, 机器学习, 预测模型

Abstract: Objective  This study aimed to compare different machine learning algorithms to identify the optimal prediction model for the short-term mortality risk in patients undergoing maintenance hemodialysis (MHD), for uremia, providing a basis for early clinical risk stratification and intervention. Methods  A total of 197 uremic patients undergoing MHD at The Second Affiliated Hospital of Air Force Military Medical University from April 2022 to May 2024 were retrospectively enrolled. Based on 12-month survival outcomes, they were categorized into a death group (n=54) and a survival group (n=143). Using independent risk factors identified by multivariate logistic regression, five machine learning models—Logistic Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN)—were developed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, F1 score, and accuracy. Decision Curve Analysis (DCA) was used to assesse clinical net benefit, and SHapley Additive exPlanations (SHAP) was applied to quantify the contribution of each clinical feature to the predictions. Results  Multivariate Logistic regression showed that diabetes (OR=14.347, 95%CI: 1.796～114.624, P=0.012), total cholesterol (OR=62.811, 95%CI: 2.463～1601.852, P=0.012), CRP (OR=4.723, 95% CI: 1.592-14.011, P=0.005) were independent risk factors for short-term mortality in patients with uremic MHD (P <0.05), while ALB (OR=-0.917, 95%CI: (0.231～0.693, P=0.001) was an independent protective factors (P<0.05). Comparison of the performance of the five algorithms reveals that the AUC value of the KNN model achieved the highest AUC value of 0.909, and an F1 score of 0.796, identifying it as the optimal predictive model for short-term mortality risk in uremic MHD patients. SHAP analysis identified ALB, CRP, diabetes, and total cholesterol as the most influential predictors. Conclusion: The KNN model integrates key indicators such as ALB, CRP, diabetes and total cholesterol, effectively predicts the short-term mortality risk in patients with uremic MHD.

Key words: Maintenance hemodialysis, Death, Machine learning, Predictive modeling

中图分类号: