Supplementary Material for: Improved Mortality Prediction in Dialysis Patients Using Specific Clinical and Laboratory Data

<b><i>Background:</i></b> Risk prediction models can be used to inform patients undergoing renal replacement therapy about their survival chances. Easily available predictors such as registry data are most convenient, but their predictive value may be limited. We aimed to improve a simple prediction model based on registry data by incrementally adding sets of clinical and laboratory variables. <b><i>Methods:</i></b> Our data set includes 1,835 Dutch patients from the Netherlands Cooperative Study on the Adequacy of Dialysis. The potential survival predictors were categorized on availability. The first category includes easily available clinical data. The second set includes laboratory values like albumin. The most laborious category contains glomerular filtration rate (GFR) and Kt/V. Missing values were substituted using multiple imputation. Within 1,225 patients, we recalibrated the registry model and subsequently added parameter sets using multivariate Cox regression analyses with backward selection. On the other 610 patients, calibration and discrimination (C-index, integrated discrimination improvement (IDI) index and net reclassification improvement (NRI) index) were assessed for all models. <b><i>Results:</i></b> The recalibrated registry model showed adequate calibration and discrimination (C-index = 0.724). Adding easily available parameters resulted in a model with 10 predictors, with similar calibration and improved discrimination (C-index = 0.784). The IDI and NRI indices confirmed this, especially for short-term survival. Adding laboratory values resulted in an alternative model with similar discrimination (C-index = 0.788), and only the NRI index showed minor improvement. Adding GFR and Kt/V as candidate predictors did not result in a different model. <b><i>Conclusion:</i></b> A simple model based on registry data was enhanced by adding easily available clinical parameters.

<b><i>背景：</i></b> 风险预测模型可用于为接受肾脏替代治疗（renal replacement therapy）的患者提供其生存概率相关信息。诸如登记数据（registry data）这类易于获取的预测因子最为便捷，但其预测价值或存在一定局限。本研究旨在通过逐步加入临床与实验室变量集，优化一款基于登记数据的简易预测模型。<b><i>方法：</i></b> 本数据集包含来自荷兰透析充分性合作研究（Netherlands Cooperative Study on the Adequacy of Dialysis）的1835名荷兰患者。潜在生存预测因子按可获取程度进行分类：第一类为易于获取的临床数据；第二类为白蛋白等实验室检测指标；最为耗时的类别则包含肾小球滤过率（glomerular filtration rate, GFR）与Kt/V。采用多重插补（multiple imputation）对缺失值进行补全。在1225名患者中，我们对登记数据模型进行了重新校准，随后通过带向后选择法的多变量Cox回归分析（multivariate Cox regression analyses）加入参数集。在剩余的610名患者中，我们对所有模型的校准性能与区分性能（C指数（C-index）、综合区分改善指数（integrated discrimination improvement, IDI）与净重新分类改善指数（net reclassification improvement, NRI））进行了评估。<b><i>结果：</i></b> 重新校准后的登记数据模型展现出良好的校准性能与区分性能（C指数=0.724）。加入易于获取的参数后，得到包含10个预测因子的模型，其校准性能与原模型相当，但区分性能得到提升（C指数=0.784）。综合区分改善指数与净重新分类改善指数验证了这一结果，尤其在短期生存预测中更为显著。加入实验室检测指标后，得到另一款区分性能相近的模型（C指数=0.788），仅净重新分类改善指数呈现小幅提升。加入肾小球滤过率与Kt/V作为候选预测因子后，并未得到性能不同的模型。<b><i>结论：</i></b> 基于登记数据的简易模型可通过添加易于获取的临床参数得到优化。