Abstract:

Objective　To develop a predictive model for intradialytic hypotension (IDH) using machine learning techniques. Methods　A retrospective analysis was conducted on the demographic data and dialysis records of the patients who underwent hemodialysis at Fuding City Hospital between October 2020 and August 2022. The variables included age, gender, pre-dialysis blood pressure, and pre-dialysis weight. Three distinct machine learning algorithms, light Gradient Boosting Machine (LGBM), support vector machine (SVM), and TabNet, were employed to construct two predictive models, designated as IDH-1 and IDH-2. The IDH-1 model integrates real-time pre-dialysis data with historical dialysis data averages to predict IDH risk instantaneously. Conversely, the IDH-2 model incorporates comprehensive current dialysis data along with historical averages to forecast IDH risk during the subsequent dialysis session. The areas under the curves (AUC), accurate rates, and F1 scores by the three algorithms were compared. Results　A total of 77 808 hemodialysis treatment records of 434 patients were used as the initial data set. After rigorous data screening, the final data set of the IDH-1 model contained 416 patients and 71 427 hemodialysis records, and the IDH-2 model contained 416 patients and 71 011 hemodialysis records. TabNet outperformed both LGBM and SVM. The AUC of the TabNet algorithm in the IDH-1 model was 0.84, with a 95% confidence interval (CI) ranging from 0.810 to 0.860. In the IDH-2 model, the AUC of the TabNet algorithm was 0.83, with a 95%CI ranging from 0.805 to 0.850. Historical frequency of IDH episodes, as well as pre-dialysis and intra-dialysis systolic blood pressures, were identified as critical predictive factors for IDH. Conclusions　This study underscores the significant potential of employing machine learning methodologies, in conjunction with demographic data and dialysis parameters, to predict IDH in hemodialysis patients.

Key words:

Chronic kidney disease, Among  , them,TabNet  , has  , the  , best  , Performance  , Hemodialysis, Machine learning, Intradialytic hypotension,  , Predictive model

摘要：

目的　通过机器学习技术开发一种能预测透析内低血压（IDH）的模型。方法　回顾性分析2020年10月至2022年8月期间在福鼎市医院接受血液透析的患者人口统计学资料和透析记录，包括年龄、性别、透析前血压、透析前体重等，采用了3种不同的机器学习算法——光梯度增强机（LGBM）、支持向量机（SVM）和TabNet，构建两个预测模型，分别命名为IDH-1和IDH-2。IDH-1模型通过整合患者透析前数据与历史透析数据的平均值来实时预测IDH风险；IDH-2模型则结合患者当前透析的全部数据及历史平均值，预测其下一次透析时IDH的发生风险。比较3种算法模型在曲线下面积（AUC）、精确率、召回率和F1分数等指标上的性能。结果　434名患者共77 808例次的血液透析治疗记录作为初始数据集，经过严格的数据筛选，IDH-1模型的最终数据集包含416名患者和71 427条血液透析记录，IDH-2模型包含416名患者和71 011条血液透析记录。TabNet在性能方面优于LGBM和SVM。在IDH-1模型中，TabNet算法的AUC值为0.84，95%CI为0.810～0.860；在IDH-2模型中，TabNet算法的AUC值为0.83，95%CI为0.805～0.850。历史IDH发作频率及透析前和透析期间的收缩血压被识别为IDH的关键预测因素。结论　机器学习方法结合人口统计数据和透析参数在预测血液透析患者IDH方面具有巨大潜力，其中TabNet性能最优。

关键词:

慢性肾脏病, 血液透析, 机器学习, 透析内低血压, 预测模型