Tran Michelle, Xu Chun Anna, Wilson Jonathan, Ephraim Patti L, Shafi Tariq, Weiner Daniel E, Goldstein Benjamin A, Scialla Julia J
Department of Medicine, University of Virginia School of Medicine, Charlottesville, Virginia.
Department of Biostatistics and Bioinformatics, Duke University School of Medicine, Durham, North Carolina.
Kidney360. 2025 Mar 1;6(3):432-440. doi: 10.34067/KID.0000000681. Epub 2024 Dec 16.
We found poor ability to distinguish among different causes of death using clinical information in the 30 days before death for patients on hemodialysis. We found that models of different causes of death identified similar associated clinical factors. Given the lack of clear phenotypic differences, this study questions the usefulness of cause of death in research in dialysis.
Patients treated with maintenance hemodialysis are at high risk of death from a variety of causes.
To identify markers (., risk phenotypes) that distinguish among causes of death, we used dialysis electronic health record data for a cohort of adults treated with maintenance in-center hemodialysis who died between 2003 and 2016 (=19,793). Patients were linked to the United States Renal Data System (USRDS) Files. We classified USRDS-reported causes of death into five categories: sudden cardiac death, nonsudden cardiac death cardiovascular death, infection, others, and unknown. A subcohort was linked to the National Death Index with similar categories defined. We used ensemble classification trees to discriminate among causes using demographics, vital signs, laboratory measures, health service utilization, and comorbidity claims from 30 days before death. We then created nested case-control populations for each cause of death and used ridge logistic regression to evaluate clinical risk markers that associate with distinct causes.
The area under the receiver operating characteristic curves from ensemble classification trees were all between 0.59 and 0.70, suggesting minimal ability to distinguish among causes using clinical risk markers. Model coefficients were similar and highly correlated across different cause of death models (., 0.87–0.94). This suggests that most clinical risk markers are shared across causes without distinct risk phenotypes.
We conclude that different causes of death may share similar clinical risk markers in the setting of kidney failure or that the causes of death attributed on USRDS or National Death Index forms are not precise.
我们发现，对于接受血液透析的患者，利用死亡前30天的临床信息来区分不同死因的能力较差。我们发现，不同死因的模型识别出了相似的相关临床因素。鉴于缺乏明确的表型差异，本研究对死因在透析研究中的实用性提出了质疑。
接受维持性血液透析治疗的患者面临多种原因导致的高死亡风险。
为了识别区分死因的标志物（如风险表型），我们使用了2003年至2016年间死亡的接受维持性中心血液透析治疗的成年队列的透析电子健康记录数据（n = 19,793）。患者与美国肾脏数据系统（USRDS）文件相关联。我们将USRDS报告的死因分为五类：心源性猝死、非心源性猝死、心血管死亡、感染、其他和不明。一个亚队列与国家死亡指数相关联，定义了相似的类别。我们使用集成分类树，根据死亡前30天的人口统计学、生命体征、实验室检查、医疗服务利用和合并症索赔来区分死因。然后，我们为每种死因创建了嵌套病例对照人群，并使用岭逻辑回归来评估与不同死因相关的临床风险标志物。
集成分类树的受试者操作特征曲线下面积均在0.59至0.70之间，表明利用临床风险标志物区分死因的能力有限。不同死因模型的模型系数相似且高度相关（如，0.87 - 0.94）。这表明大多数临床风险标志物在不同死因之间共享，没有明显的风险表型。
我们得出结论，在肾衰竭情况下，不同死因可能共享相似的临床风险标志物，或者USRDS或国家死亡指数表格上归因的死因并不准确。
