Survival prognosis after the start of a renal replacement therapy in the Netherlands: a retrospective cohort study
© Hemke et al.; licensee BioMed Central Ltd. 2013
Received: 29 January 2013
Accepted: 6 November 2013
Published: 20 November 2013
There is no single model available to predict the long term survival for patients starting renal replacement therapy (RRT). The available models either predict survival on dialysis until transplantation, survival on the transplant waiting list, or survival after transplantation. The aim of this study was to develop a model that includes dialysis survival and survival after an eventual transplantation.
From the Dutch renal replacement registry, patients of 16 years of age or older were included if they started RRT between 1995 and 2005, still underwent RRT at baseline (90 days after the start of RRT) and were not registered at a non-renal organ transplant waiting list (N = 13868). A prediction model of 10-year patient survival after baseline was developed through multivariate Cox regression analysis, in one half of the research group. Age at start, sex, primary renal disease (PRD) and therapy at baseline were included as possible predictors. A sensitivity analysis has been performed to determine whether listing on the transplant waiting list should be added. The predictive performance of the model was internally validated. Calibration and discrimination were computed in the other half of the research group. Another sensitivity analysis was to assess whether the outcomes differed if the model was developed and tested in two geographical regions, which were less similar than the original development and validation group. No external validation has been performed.
Survival probabilities were influenced by age, sex, PRD and therapy at baseline (p < 0.001). The calibration and discrimination both showed very reasonable results for the prediction model (C-index = 0.720 and calibration slope for the prognostic index = 1.025, for the 10 year survival). Adding registration on the waiting list for renal transplantation as a predictor did not improve the discriminative power of the model and was therefore not included in the model.
With the presented prediction model, it is possible to give a reasonably accurate estimation on the survival chances of patients who start with RRT, using a limited set of easily available data.
In the Netherlands, in recent years approximately 2000 new patients with end stage renal disease (ESRD) start chronic renal replacement therapy (RRT) every year. Even though the kidney replacement programs already exist for more than forty years, it is still not possible to predict the long term survival chances for all RRT patients during the initial phase of their therapy, using one single model.
Existing prediction models look at dialysis survival until transplantation , patient survival on the transplant waiting list [2–5], patient survival after transplantation [6, 7], or focus on a specific patient group in which differences in treatment modality are less likely . However, none of the available predication models focus on survival for the complete group of incident RRT patients, taking into account survival after dialysis combined with survival after a possible transplantation. As it is not clear at therapy initiation whether a patient will stay on dialysis, or will be listed in time and actually be transplanted, the available models cannot be used to predict survival for all patients at the start of RRT.
To be able to give a survival prognosis in an early stage of the renal replacement therapy to every patient, we need a model that predicts patient survival chances based on characteristics that are known at that point in time. In the present study, based on national data from the renal replacement registry, a prediction model on the survival prognosis for incident RRT patients in the Netherlands was developed and validated.
The objective of this study is to develop a prediction model that could be used by physicians to inform patients about their survival chances at the start of RRT, based on a few very easily obtainable variables.
Demographics of patients, N = 13868
Patients starting a renal replacement therapy in 1995–2005, ≥16 years of age, with a registered primary renal disease and peritoneal dialysis, hemodialysis or a functioning kidney transplant at 90 days after the start
Total N = 13868 %
Development group N = 6934 %
Validation group N = 6934 %
75 year or older
Primary renal disease
Renal vascular disease
Cystic kidney disease
Start year renal replacement
1995 – 2000
2001 – 2005
Therapy at baseline
The analysis was performed using SPSS 19 and STATA. Survival was analyzed with Kaplan Meier and log rank tests. Linearity of the influence of patient age on survival was assessed with Kaplan Meier stratified by different age groups. The proportionality assumption has been tested by visual inspection of the Schoenfeld residuals plot . For the survival prognosis from 90 days after the start of RRT, multivariate Cox regression analysis was performed in the development group. The formula for the survival probability at time t, S(t), is S(t) = exp(-H(t)). Here H(t) is the cumulative hazard that is calculated from the baseline hazard (H0) as H(t) = H0(t)*exp(prognostic index). The prognostic index (PI) is the sum of the parameter estimates from the Cox regression multiplied by the patient characteristics for a specific patient. To validate the prediction model, the predictive performance was assessed by computing the calibration and the discrimination of the prediction model for the 3, 5 and 10 year survival in the validation group [10, 11]. Calibration refers to the agreement between observed outcomes and predicted survival probabilities. This was measured by a) the calibration in the large, which indicates the extent that predictions are systematically too low or too high, b) a calibration plot for ten deciles according to the predicted survival, which is plotted against the observed survival, which ideally should be on the 45-degree line, and c) the calibration slope, which should be 1. The discrimination is the ability of the model to distinguish subjects with different outcomes. This was measured by the concordance (or C-) index. A C-index of 0.5 indicates the model has no discriminative power, while a model with a C-index of 1.0 has a perfect discriminative power. As a sensitivity analysis we assessed the consequences of our choice not to include information about registration on the waiting list as one of the predictors in the model and the choice for random (instead of geographical) development and validation group stratification.
Cox regression model for patient mortality 90 days after start of renal replacement therapy
95% confidence interval
Primary renal disease
Cystic kidney disease
Renal vascular disease
Therapy at 90 days
Age (per year)
Baseline hazards for the referent patient group
To assess the predictive performance of the model, the calibration and discrimination were computed in the validation group. The calibration in the large, or overall calibration, was good with a 50.4% predicted versus 49.5% observed 5-year survival and 32.9% predicted versus 34.4% observed 10-year survival.
To show robustness of our model, we also performed sensitivity analyses.
As a first sensitivity analysis the model has been extended with the registration on the kidney waiting list at 90 days, combined with the therapy at 90 days and presented together in the model as the status at 90 days. This model had a similar validity to our current model with a discrimination of 0.724 and a calibration in the large equal, and a calibration by deciles almost equal, to that of the current model.
As the development group and validation group were very similar, we also assessed the influence of dividing the research cohort into two geographical regions, based on the ZIP-codes of the patients’ addresses. The two resulting comparable sized regions differed from each other in age-distribution, PRD-distribution, starting period, therapy at 90 days and transplantation rate. One of these regions also differed from the development group on all mentioned items; the other only differed in PRD-distribution. Two sensitivity analyses have been performed. The first was the development of a model in one region and validating the outcome in the other region. Parameter estimates of this alternative model did not differ substantially from our original (and final) model. This model had a similar validity to our final model with a discrimination of 0.711 and a calibration almost equal to that of the final model. The final model was also validated in the two separate regions. The model performed well in both regions, with a similar discrimination (C = 0.71) and calibration slopes of the prognostic index of 0.982 and 1.040 respectively (data not shown).
A prediction model was developed to estimate survival probabilities at 90 days based on a basic set of patient characteristics (age at start of RRT, sex and PRD) and the RRT therapy at 90 days. The main strength of the current prediction model is that it is based on the complete cohort of Dutch patients in 1995–2005. The predictive performance of the model is adequate, as demonstrated by validity tests on calibration and discrimination of this model, which could thus be used to inform patients about their survival prognosis at baseline.
The model uses treatment information at 90 days after the start of RRT, as the intention to treat. There is a clear difference in survival between patients who are on dialysis or who are being transplanted in an early stage. From previous studies we know that a better survival for kidney transplantation can be related to both advantages of the therapy as well as the better condition of the patient [2, 12]. There is also a survival difference between patients starting on hemodialysis and peritoneal dialysis, which is also known from literature on this topic. Research has shown that patients starting on hemodialysis have more co-morbidities than patients starting on peritoneal dialysis [13–15]. In the prognostic formula the therapy modality therefore is included as one of the indicators of patient condition, as no other clinical information is available in the complete Dutch patient cohort. Like the ERA-EDTA, the Dutch renal replacement registry only collects a few parameters on all patients. Further research should establish whether the treatment at 90 days, in combination with age and PRD, is a good alternative for clinical parameters indicating the patient’s condition.
Registration on the kidney transplant waiting list at 90 days was not included as one of the predictors. Other studies have shown the survival benefit of patients registered at the waiting list, compared to dialysis patients not listed for transplantation [12, 13, 16], suggesting that this predictor is related to patient condition. For that reason we performed a sensitivity analysis to test the possible additional value of this predictor. The additional predictive performance, however, was negligible. Possibly this could indicate that there is an overlapping risk profile between dialysis patients listed and not listed for transplantation as has been shown by an American study , where they found that many ESRD patients viable for transplantation were not listed while higher risk patients had been listed rapidly. The comparable performance of the prediction models with and without registration on 90 days suggests that the chance to be in a better general condition and/or to be transplanted is not only reflected in registration on the waiting list, but is also covered by the other predictors (age at start, PRD and therapy at 90 days). Another reason not to include registration on the kidney transplant waiting list as a predictor in our model is the fact that the time point of registration is very arbitrary in the Netherlands. There was a very large variation in registration time in Dutch population, and many patients were registered on the waiting list for kidney transplantation after the period of 90 days, which was our baseline for inclusion. It is not very likely that there is a difference in condition between patients that are registered at 90 days or, for instance, at 91 days after the start of RRT.
Some potential limitations of this study should also be noted.
First, the moderate discriminative power of the prediction model (C-index of 0.720) shows further improvement possibilities. In this study the age, PRD, and therapy at 90 day are considered to be substitutes for more accurate clinical indicators on the condition of patients. Adding clinical patient characteristics, like GFR, proteinuria, and (historical) co-morbidities, would probably improve the individual prediction. An English study showed that the addition of comorbid condition data and laboratory data could indeed lead to improvement of the predictive power of the prognostic model from 0.69 to 0.75 . On the other hand the additional effect may be limited, as age and PRD are highly correlated with comorbidity, as has also be shown by a European study  and a single centre study in the US . It is therefore desirable to study for the Dutch situation whether clinical data correlate with data already used in this model and whether they can improve the discriminative power of the prediction model.
Another limitation of the study is that the model is only internally validated in the validation group, and the model has not been externally validated in an external cohort. It would be desirable to test the model in another patient group. This could be a patient cohort from another country or another period. The generalizability of the model to another country, however, is doubtful, as countries differ in dialysis and transplantation possibilities. This should be subject for further research. The fact that our model focuses on long term survival, makes external validation in a more recent cohort difficult. Regular evaluation of the model is needed as treatments improve in time and RRT-population, treatment possibilities and choices, both in dialysis and transplantation, change.
Finally, note that the prediction model presented in this study can only be used to inform patients about their survival chances from 90 days after their start of RRT. The patients for whom the model can be used, should have survived the first 90 days of RRT and the therapy choice for their RRT therapy has been made earlier. This prediction model is not suitable to be used for the choice of the therapy modality at the start of the RRT or for the acceptance or decline of a specific transplant kidney offer. The therapy choice should be based on preferences of the patient and physician, as is also the case for the choice to accept or decline a specific transplant kidney offer. For the choices between therapies and the probability of death on therapies new designs are currently emerging, based on competing risks instead of the Kaplan Meier method .
In conclusion, with the presented prediction model it is possible to give a reasonably accurate estimation on the survival chances of patients who start with RRT, using a limited set of easily available data. Future research should establish whether it is possible to improve the predictive performance of the prediction model using more clinical parameters.
We would like to thank all centres participating in the renal replacement registry in the Netherlands and the Dutch organ transplant registration.
- Wagner M, Ansell D, Kent DM, Griffith JL, Naimark D, Wanner C, Tangri N: Predicting mortality in incident dialysis patients: an analysis of the United Kingdom Renal Registry. Am J Kidney Dis. 2011, 57: 894-902. 10.1053/j.ajkd.2010.12.023.View ArticlePubMedPubMed CentralGoogle Scholar
- van Walraven C, Austin PC, Knoll G: Predicting potential survival benefit of renal transplantation in patients with chronic kidney disease. CMAJ. 2010, 182: 666-672. 10.1503/cmaj.091661.View ArticlePubMedPubMed CentralGoogle Scholar
- Rabbat CG, Thorpe KE, Russell JD, Churchill DN: Comparison of mortality risk for dialysis patients and cadaveric first renal transplant recipients in Ontario, Canada. J Am Soc Nephrol. 2000, 11: 917-922.PubMedGoogle Scholar
- Medin C, Elinder CG, Hylander B, Blom B, Wilczek H: Survival of patients who have been on a waiting list for renal transplantation. Nephrol Dial Transplant. 2000, 15: 701-704. 10.1093/ndt/15.5.701.View ArticlePubMedGoogle Scholar
- Schold JD, Meier-Kriesche HU: Which renal transplant candidates should accept marginal kidneys in exchange for a shorter waiting time on dialysis?. Clin J Am Soc Nephrol. 2006, 1: 532-538. 10.2215/CJN.01130905.View ArticlePubMedGoogle Scholar
- Jassal SV, Schaubel DE, Fenton SS: Predicting mortality after kidney transplantation: a clinical tool. Transpl Int. 2005, 18: 1248-1257. 10.1111/j.1432-2277.2005.00212.x.View ArticlePubMedGoogle Scholar
- Kasiske BL, Israni AK, Snyder JJ, Skeans MA, Peng Y, Weinhandl ED: A simple tool to predict outcomes after kidney transplant. Am J Kidney Dis. 2010, 56: 947-960. 10.1053/j.ajkd.2010.06.020.View ArticlePubMedGoogle Scholar
- Couchoud C, Labeeuw M, Moranne O, Allot V, Esnault V, Frimat L, Stengel B: A clinical score to predict 6-month prognosis in elderly patients starting dialysis for end-stage renal disease. Nephrol Dial Transplant. 2009, 24: 1553-1561. 10.1093/ndt/gfn698.View ArticlePubMedGoogle Scholar
- Schoenfeld D: Partial Residuals for The Proportional Hazards Regression Model. Biometrika. 1982, 69: 239-241. 10.1093/biomet/69.1.239.View ArticleGoogle Scholar
- Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, Pencina MJ, Kattan MW: Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology. 2010, 21: 128-138. 10.1097/EDE.0b013e3181c30fb2.View ArticlePubMedPubMed CentralGoogle Scholar
- Vergouwe Y, Steyerberg EW, Eijkemans MJC, Habbema JDF: Validity of prediction models: when is a model clinically useful?. Semin Urol Oncol. 2002, 20: 96-107. 10.1053/suro.2002.32521.View ArticlePubMedGoogle Scholar
- Wolfe RA, Ashby VB, Milford EL, Ojo AO, Ettenger RE, Agodoa LY, Held PJ, Port FK: Comparison of mortality in all patients on dialysis, patients on dialysis awaiting transplantation, and recipients of a first cadaveric transplant. N Engl J Med. 1999, 341: 1725-1730. 10.1056/NEJM199912023412303.View ArticlePubMedGoogle Scholar
- Termorshuizen F, Korevaar JC, Dekker FW, Van Manen JG, Boeschoten EW, Krediet RT: Hemodialysis and peritoneal dialysis: comparison of adjusted mortality rates according to the duration of dialysis: analysis of The Netherlands Cooperative Study on the Adequacy of Dialysis 2. J Am Soc Nephrol. 2003, 14: 2851-2860. 10.1097/01.ASN.0000091585.45723.9E.View ArticlePubMedGoogle Scholar
- Liem YS, Wong JB, Hunink MG, de Charro FT, Winkelmayer WC: Comparison of hemodialysis and peritoneal dialysis survival in The Netherlands. Kidney Int. 2007, 71: 153-158. 10.1038/sj.ki.5002014.View ArticlePubMedGoogle Scholar
- Vonesh EF, Snyder JJ, Foley RN, Collins AJ: The differential impact of risk factors on mortality in hemodialysis and peritoneal dialysis. Kidney Int. 2004, 66: 2389-2401. 10.1111/j.1523-1755.2004.66028.x.View ArticlePubMedGoogle Scholar
- Heaf JG, Lokkegaard H, Madsen M: Initial survival advantage of peritoneal dialysis relative to haemodialysis. Nephrol Dial Transplant. 2002, 17: 112-117.View ArticlePubMedGoogle Scholar
- Schold JD, Srinivas TR, Kayler LK, Meier-Kriesche HU: The overlapping risk profile between dialysis patients listed and not listed for renal transplantation. Am J Transplant. 2008, 8: 58-68.View ArticlePubMedGoogle Scholar
- van Manen JG, van Dijk PC, Stel VS, Dekker FW, Cleries M, Conte F, Feest T, Kramar R, Leivestad T, Briggs JD, Stengel B, Jager KJ: Confounding effect of comorbidity in survival studies in patients on renal replacement therapy. Nephrol Dial Transplant. 2007, 22: 187-195.View ArticlePubMedGoogle Scholar
- Beddhu S, Bruns FJ, Saul M, Seddon P, Zeidel ML: A simple comorbidity scale predicts clinical outcomes and costs in dialysis patients. Am J Med. 2000, 108: 609-613. 10.1016/S0002-9343(00)00371-5.View ArticlePubMedGoogle Scholar
- Beuscart JB, Pagniez D, Boulanger E, Lessore de Sainte FC, Salleron J, Frimat L, Duhamel A: Overestimation of the probability of death on peritoneal dialysis by the Kaplan-Meier method: advantages of a competing risks approach. BMC Nephrol. 2012, 13: 31-10.1186/1471-2369-13-31.View ArticlePubMedPubMed CentralGoogle Scholar
- The pre-publication history for this paper can be accessed here:http://www.biomedcentral.com/1471-2369/14/258/prepub
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.