Geddes Colin C, van Dijk Paul C W, McArthur Stephen, Metcalfe Wendy, Jager Kitty J, Zwinderman Aeilko H, Mooney Michael, Fox Jonathan G, Simpson Keith
Renal Unit, Western Infirmary, Dumbarton Road, Glasgow G11 6NT, UK.
Nephrol Dial Transplant. 2006 Apr;21(4):945-56. doi: 10.1093/ndt/gfi326. Epub 2005 Dec 8.
Accurate prediction of patient survival from the time of starting renal replacement therapy (RRT) is desirable, but previously published predictive models have low accuracy. We have attempted to overcome limitations of previous studies by conducting an ambidirectional inception cohort study in patients on RRT from centres throughout Europe. A conventional multivariate regression (MVR) model, a self-learning rule-based model (RBM) and a simple co-morbidity score [the Charlson score modified for renal disease (MCS)] were compared.
In 1996, all 3640 dialysis centres registered with the ERA-EDTA were invited to identify all patients on RRT for end-stage renal failure (ESRF) who died during the 28 days of February 1997 (training cohort) and all patients who started RRT in the same period (validation cohort). Fifty-four clinical and laboratory variables from the time of starting RRT were collected in both cohorts using a two-page questionnaire. The data from the training cohort were given to statisticians at the Amsterdam Academic Medical Centre to create the MVR model and to engineers in Strathclyde University to create the RBM. They were then given the baseline data from patients in the validation cohort to predict how long each patient would survive. Follow-up questionnaires were sent to the centre of each patient in the validation cohort to determine actual survival.
A total of 2310 patients from 793 centres in 37 countries in the ERA-EDTA area were used to construct and validate the models. For predicting 1-year survival, the RBM had the highest positive predictive value (PPV) (84.2%), the MVR model had the highest negative predictive value (NPV) (47%) and the RBM had the highest likelihood ratio (1.59). For predicting 5-year survival, the MCS had the highest PPV (79.4%), the RBM had the highest NPV (74.3%) and the MCS had the highest likelihood ratio (7.0). The proportion of explained variance in survival for MCS, MVR and RBM was 14.6, 12.9 and 3.95%, respectively.
Using the ambidirectional inception cohort design of this ERA-EDTA Registry survey, we have been able to create and validate two novel instruments to predict survival in patients starting RRT and compare them with a simple scoring model. The models tended to predict 5-year survival with more accuracy than 1-year survival. Examples of potential applications include informing clinical decision making about the likely benefit of starting RRT and listing for transplantation, adjusting for baseline risk in comparative studies and identifying specific risk groups to participate in clinical trials.
从开始肾脏替代治疗（RRT）时准确预测患者生存率是很有必要的，但此前发表的预测模型准确性较低。我们试图通过在来自欧洲各地中心的接受RRT治疗的患者中开展一项双向起始队列研究来克服以往研究的局限性。对传统多变量回归（MVR）模型、基于自学习规则的模型（RBM）和一个简单的合并症评分[针对肾脏疾病修正的查尔森评分（MCS）]进行了比较。
1996年，邀请了向欧洲肾脏协会 - 欧洲透析与移植协会（ERA - EDTA）注册的所有3640个透析中心，确定在1997年2月28天内死亡的所有接受RRT治疗的终末期肾衰竭（ESRF）患者（训练队列）以及同期开始RRT治疗的所有患者（验证队列）。使用一份两页的问卷收集了两个队列中从开始RRT时起的54个临床和实验室变量。训练队列的数据被提供给阿姆斯特丹学术医疗中心的统计学家以创建MVR模型，并提供给斯特拉斯克莱德大学的工程师以创建RBM。然后将验证队列中患者的基线数据提供给他们，以预测每位患者的生存时长。向验证队列中每位患者的中心发送随访问卷以确定实际生存情况。
来自ERA - EDTA地区37个国家793个中心的2310名患者被用于构建和验证模型。对于预测1年生存率，RBM具有最高的阳性预测值（PPV）（84.2%），MVR模型具有最高的阴性预测值（NPV）（47%），且RBM具有最高的似然比（1.59）。对于预测5年生存率，MCS具有最高的PPV（79.4%），RBM具有最高的NPV（74.3%），且MCS具有最高的似然比（7.0）。MCS、MVR和RBM对生存率的解释方差比例分别为14.6%、12.9%和3.95%。
利用ERA - EDTA注册研究的双向起始队列设计，我们得以创建并验证了两种预测开始RRT治疗患者生存率的新工具，并将它们与一个简单的评分模型进行比较。这些模型预测5年生存率的准确性往往高于1年生存率。潜在应用的例子包括为关于开始RRT治疗可能益处的临床决策提供信息以及列入移植名单、在比较研究中调整基线风险以及识别参与临床试验的特定风险组。
