Few risk scores are available for predicting mortality in chronic kidney disease (CKD) patients undergoing predialysis nephrology care. Here, we developed a risk score using predialysis nephrology practice data to predict 1-year mortality following the initiation of haemodialysis (HD) for CKD patients.
This was a multicenter cohort study involving CKD patients who started HD between April 2006 and March 2011 at 21 institutions with nephrology care services. Patients who had not received predialysis nephrology care at an estimated glomerular filtration rate (eGFR) of approximately 10 mL/min per 1.73 m2 were excluded. Twenty-nine candidate predictors were selected, and the final model for 1-year mortality was developed via multivariate logistic regression and was internally validated by a bootstrapping technique.
A total of 688 patients were enrolled, and 62 (9.0%) patients died within one year of HD initiation. The following variables were retained in the final model: eGFR, serum albumin, calcium, Charlson Comorbidity Index excluding diabetes and renal disease (modified CCI), performance status (PS), and usage of erythropoiesis-stimulating agent (ESA). Their β-coefficients were transformed into integer scores: three points were assigned to modified CCI≥3 and PS 3–4; two to calcium>8.5 mg/dL, modified CCI 1–2, and no use of ESA; and one to albumin<3.5 g/dL, eGFR>7 mL/min per 1.73 m2, and PS 1–2. Predicted 1-year mortality risk was 2.5% (score 0–4), 5.5% (score 5–6), 15.2% (score 7–8), and 28.9% (score 9–12). The area under the receiver operating characteristic curve was 0.83 (95% confidence interval, 0.79–0.89).
Citation: Doi T, Yamamoto S, Morinaga T, Sada K-e, Kurita N, Onishi Y (2015) Risk Score to Predict 1-Year Mortality after Haemodialysis Initiation in Patients with Stage 5 Chronic Kidney Disease under Predialysis Nephrology Care. PLoS ONE 10(6): e0129180. https://doi.org/10.1371/journal.pone.0129180
Academic Editor: Tatsuo Shimosawa, The University of Tokyo, JAPAN
Received: October 20, 2014; Accepted: May 5, 2015; Published: June 9, 2015
Copyright: © 2015 Doi et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited
Data Availability: All relevant data are within the paper.
Funding: This work was supported by grants from the Institute for Health Outcomes & Process Evaluation research. The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Although intervention by nephrologists to chronic kidney disease (CKD) patients before initiation of haemodialysis (HD) is increasing, these patients still have poor outcome. Therefore, it is important for the nephrologists to practice shared decision making with CKD patients and their families regarding the initiation of HD based on informed estimates of early mortality risk. Although several risk scores based on the clinical parameters at the initiation of HD have been previously reported, the risk score to predict mortality risk after initiation of dialysis treatment among CKD patients under predialysis nephrology care is scarce.
Couchoud et al. developed a simple clinical score predicting early mortality in CKD patients starting HD . Although this score was a simple integer score, this study only enrolled elderly patients and included those with unplanned initiation of HD. In addition, the estimated glomerular filtration rate (eGFR) at the initiation of HD was not incorporated into this score, despite being a strong prognostic factor  that varies by country and should therefore be considered in mortality prediction. For example, the proportion of patients with eGFR>10 mL/min at the initiation of HD was 11% in Japan  and 36% in USA . van Diepen et al. developed a predictive model for incident dialysis patients with consideration of residual GFR as a candidate predictor, whereas this model was only applicable to diabetic dialysis patients and difficult to calculate . Therefore, a simple risk score using daily practice data is necessary for nephrologists to predict mortality of CKD patients after initiation of dialysis treatment without restriction of patients’ age or underlying diseases.
Using multicenter clinical data measured immediately before the initiation of HD, we developed and validated a clinical scoring system predicting 1-year survival after initiation of HD for CKD patients.
Subjects and Methods
We conducted a retrospective cohort study of patients starting HD at 19 tertiary care institutions and two clinics. This study was approved by the ethics committee of the Okayama University Graduate School of Medicine, Dentistry and Pharmaceutical Sciences (authorization number: 538) and registered with the University Hospital Medical Information Network Clinical Trials Registry (UMIN000007862). Informed consent from participants was ruled unnecessary because we retrospectively analyzed data routinely collected from all patients and we analyzed the data anonymously.
In this study, renal function was evaluated by eGFR using the following equation developed for Japanese patients: eGFR (mL/min per 1.73 m2) = 194 × Serum creatinine-1.094 × Age-0.287 (× 0.739 if female) . The inclusion criteria for this study were as follows: patients who started HD between April 2006 and March 2011; and eGFR during predialysis nephrology care could be confirmed in medical records and were shown to have reached to 10 mL/min per 1.73 m2 to eliminate patients with unplanned HD initiation. The criteria for exclusion from this study were as follows: those whose eGFR could not be confirmed during predialysis nephrology care because of unplanned initiation or failure to measure; those who started HD before their eGFR had reached to 10 mL/min per 1.73 m2; withdrawal of HD treatment within one year; and under 20 years old.
Data Collection and Processing
Data at the initiation of HD were collected from medical records, including demographic information, underlying renal disease, clinical symptoms, laboratory data, Charlson Comorbidity Index (CCI) , performance status categorised using scales of the World Health Organization performance status except category 5 (death) , and treatment status.
Outcome Measures & Candidate Predictors
The primary outcome measure was all-cause mortality within one year of starting HD. Candidate predictors were selected according to a literature review and clinical expertise. Candidate predictors need to be biologically plausible and available in the majority of nephrology practice settings. We therefore selected information on 29 variables, including patient demographics (age, gender, primary renal disease [diabetic nephropathy or other renal disease], and body mass index), laboratory data (eGFR, haemoglobin, urea nitrogen, serum potassium, calcium, phosphorus, albumin, and C-reactive protein), CCI, performance status, symptoms related to uremia (fatigue, edema, pulmonary edema, nausea, dysorexia, diarrhea, constipation, other digestive symptoms, hypertension, peripheral neuropathy, psychiatric disorder, haemorrhagic diathesis, diabetic retinopathy, and itch), and nephrologists’ treatments (erythropoiesis-stimulating agent [ESA]) (Table 1). Uremic symptoms are diagnosed by attending physician. In Japan, almost all patients initiating HD hand in the certificate of attending physician for making use of public support. This certificate includes the following items: chest X-ray, fundus feature, peripheral neuropathy, digestive symptom, psychiatric disorder, severe hypertension, and so on. We collected the data about uremic symptoms from these certificate and medical records. We modified CCI by excluding the items related to diabetes and renal disease in this study for the following reasons: all patients have “moderate or severe renal disease”; and diabetes is included in another variable, that is, primary renal disease.
Data Description & Handling Missing Data
Descriptive statistics on candidate predictors and the outcome variable are presented as median or percentage as appropriate.
At least one variable was missing in 183 patients (26.6%). As complete case analysis leads to loss of power and biased results, a multiple imputation approach was utilized to derive predictions for the missing values . Whereas simulation studies have shown that required number of imputations can be as three for data with 20% of missing entries , the largest percentage of missing among candidate predictors was 6.7% (46 of 688 patients) for C-reactive protein, and thus five repeated imputations would be considered a conservative choice. Five imputations were performed using multiple imputations with chained equations [10, 11]. Process of the multiple imputation and derivation of the prediction rule was shown as Fig 1.
(1) Five multiply imputed datasets were created using original data. (2) Backward elimination was separately applied to each of the five imputed datasets, resulting in five sets of selected predictors. (3) Predictors that were selected in all of the five data sets were chosen as the final set of selected predictors, with exclusion of some predictors based on balance between number of candidate predictors with number of outcomes (deaths) and discussion according to clinical relevance. (4) The logistic regression with the selected six predictors was separately applied to each of the five imputed data sets, giving five sets of β-coefficients of the six predictors. (5) To avoid overfitting, each of five sets of β-coefficients of the six predictors were shrunken using heuristic shrinkage factor. Then, the mean for each of the five estimates for β-coefficients of the final model were taken and variances of the five estimates were pooled according to Rubin’s rules. (6) The shrunken β-coefficients of the predictors in the final model divided by two-fifths of the two small β-coefficients in the model and rounded up to the nearest integer to give a simple point score.
The association of candidate predictors with 1-year mortality was modeled with multivariate logistic regression. Continuous predictor variables were categorized as appropriate with reference to β-coefficients of bivariate logistic regression models. Selection of predictors was based on backward elimination of variable with P-value>0.157, which is equivalent to the use of the Akaike information criterion (AIC) for model selection .
A majority method was selected to obtain a final set of selected predictors . First, backward elimination was separately applied to each of the five imputed data sets (Fig 1, (2)), resulting in five sets of selected predictors. Second, predictors that were selected in all of the five data sets were chosen as the final set of selected predictors (Fig 1, (3)), with exclusion of some predictors based on balance between number of candidate predictors with number of outcomes (deaths) and discussion according to clinical relevance. Third, the logistic regression with the selected predictors was separately applied to each of the five imputed data sets (Fig 1, (4)). Finally, the mean for each of the five estimates for β-coefficients of the final model were taken and variances of the five estimates were pooled according to Rubin’s rules (Fig 1, (5)) [10, 13].
In a derivation phase, overfitting should be adjusted for, especially when a large number of candidate predictors compared to number of outcomes (deaths) was considered or automated variable selection such as backward elimination was used . To account for this, the β-coefficients derived from each of the imputed data sets were multiplied by a heuristic shrinkage factor, as follows: where χ2model denotes the model chi-square and df denotes the degree of freedom (Fig 1, (5)). This approach is a post hoc method to adjust β-coefficients of the predictors for overfitting and avoids extreme predictions when applied to new patients [11, 15]. The shrunken β-coefficients of the predictors in the final model divided by two-fifths of the two small β-coefficients in the model and rounded up to the nearest integer to give a simple point score (Fig 1, (6)).
The validation was performed using single imputed data among the five imputed data generated in the derivation phase, as the single imputation validation is a temporary stand-in for the multiple imputation fit . To validate the final model, we selected a bootstrapping technique with 2000 resamples that, in contrast to the data-splitting technique, preserves the sample size and leads to more precision and power . This technique estimates the likely performance of the prediction rule on a new sample of patients from the same patient population. The ability to discriminate between patients with and without 1-year mortality was assessed by the area under the receiver operating characteristic (ROC) curve. The agreement between predicted 1-year mortality based on the rule and the observed mortality risk was assessed with a calibration plot [15, 17].
Total scores were calculated by adding the scores for each final predictor. Patients were then categorized into total score categories and their observed risks of 1-year mortality were compared.
All statistical analyses were performed using Stata version 11.2 (Stata Corp., College Station, TX, USA) and R software (http://www.R-project.org).
Description of Candidate Predictors and Outcome Variable
A total of 688 patients enrolled in this study. The candidate predictors and the outcome variables are described in Table 1. Median age of patients was 69 years, 230 patients (33.4%) were female, and 42.6% (293 of 688 patients) had diabetic nephropathy as an underlying disease. Median HD-free interval from eGFR of 10 mL/min per 1.73 m2 was 196 days, median eGFR at initiation of HD was 5.43 mL/min per 1.73 m2, and 85.8% (585 of 682 patients) were administered ESA before initiation of HD. Only 11.6% (79 of 684 patients) had a modified CCI of 0 points, and 85.3% (583 of 683 patients) had deteriorated activities of daily living defined as performance status. A total of 62 patients (9.0%) died within one year of starting HD.
Multivariable logistic regression analysis with backward elimination procedure showed that the following variables were retained in all of the five imputed dataset: eGFR>7 mL/min per 1.73 m2, serum albumin<3.5 g/dL, calcium>8.5 mg/dL, modified CCI≥1 points (higher risk for ≥3 points), performance status≥1 points (higher risk for 3–4 points), body mass index>18, pulmonary edema, and ESA use (Table 2). Among the eight candidate predictors, two were excluded: body mass index (based on discussion that body height and body weight will be unlikely to be obtained among patients with decreased performance status), pulmonary edema (based on discussion that it was anti-intuitive for future use that odds ratio of presence of pulmonary edema for death was <1.0 in all of the five imputed dataset). Further, based on number of outcomes (62 deaths), combination of the following six predictors were considered to be reasonably fitted to the final multivariable logistic regression model: eGFR>7 mL/min per 1.73 m2, serum albumin<3.5 g/dL, calcium>8.5 mg/dL, modified CCI≥1 points (higher risk for ≥3 points), performance status≥1 points (higher risk for 3–4 points), and ESA non-use (Table 3). For the analysis population, calculations showed an area under the ROC curve of 0.829.
Predictive discrimination of 1-year mortality was good, with an area under ROC curve of 0.831 (95% confidence interval, 0.789–0.892). The predicted risks of approximately 20% of the apparent model were overestimated, and the predicted risks of >40% were underestimated (Fig 2). The bias-corrected estimate was slightly nonlinear but only slightly improved compared to the apparent calibration. The 90 percentile of absolute error in predicted mortality risks between the bootstrap and original model is 0.013, which suggests only a small degree of bias from over-fitting in the original model.
The short-dashed line (“Apparent”) indicates the agreement between predicted mortality risks and observed proportions of the original model. The sold line (“Bias-corrected”) indicates the agreement between predicted mortality risks and observed proportions of the bootstrap model.
Development of Risk Score
Table 4 shows the score chart derived from the final logistic regression model (Table 3) that can be used as a risk score. A median overall risk score of six (range: 0–12) was obtained for each patient in the study population. Fig 3 shows predicted mortality risks and observed proportions for the ranges of total scores. Some discrepancies between predicted risk and observed proportion were observed (Table 5).
Prognostic score calculated form the following six items well predicts 1-year mortality for patients initiating haemodialysis: high eGFR level (>7 mL/min per 1.73 m2), low serum albumin levels, high calcium levels, high modified Charlson Comorbidity Index, low performance status, and no use of ESA. The modified Charlson Comorbidity Index was excluded items related to diabetes and renal disease from the original Charlson Comorbidity Index in the present study.
In this multicenter cohort study, we developed a simple risk score using parameters from blood examinations obtained in routine clinical evaluations to predict 1-year mortality in stage 5 CKD patients who were starting HD. This score might provide a personalized assessment of mortality risk of CKD patients under predialysis nephrology care and help nephrologists make a shared decision with patients and their families regarding the initiation of HD.
Our risk score to predict 1-year mortality is composed of the following six items available at initiation of HD: high eGFR level (>7 mL/min per 1.73 m2), low serum albumin levels, high calcium levels, high modified CCI, low performance status, and no use of ESA.
These items have a number of clinical implications or similarities to previous studies. First, despite the exclusion of eGFR patients with >10 mL/min per 1.73 m2, mortality prediction of high eGFR is in good agreement with a previous study showing that of patients who started HD, those with eGFR of 8–10 mL/min per 1.73 m2 had higher mortality than those with eGFR of 4–6 mL/min per 1.73 m2 . However, whether the initiation of HD when eGFR levels are high directly causes earlier death is still controversial. Of note, comorbidity, which causes decompensated symptoms related to the renal failure even though eGFR is relatively preserved, might exacerbate poor mortality.
Second, mortality prediction related to high modified CCI, hypoalbuminemia, and low performance status is also compatible with previous reports [5, 18, 19]. Hypoalbuminemia might indicate malnutrition, nephrotic, or chronic inflammatory state and can be partially modified by medical care. Predialysis management for nutrition and proteinuria might therefore improve survival after initiation of HD.
Third, accurate mortality prediction based on elevated serum calcium levels is consistent with previous findings of mortality risk related to abnormally elevated calcium levels among CKD patients and the incidence HD [20–25]. Increased serum calcium levels are generally caused by certain drugs, presence of malignancies or immobility. It might be speculated that one of the reasons of higher serum calcium levels in predialysis CKD patients is to use drugs such as calcium-based phosphate binder, vitamin D or thiazide as it was independent of serum albumin level and lower performance status as proxy for immobility. For example, coronary calcification induced by calcium-based phosphate binders occurred more than that by sevelamer in CKD patients .
Fourth, mortality prediction with no use of an ESA before initiation of HD is consistent with previous reports [27, 28]. However, the beneficial effect of an ESA for CKD before the initiation of HD is controversial. In addition, the use of an ESA in the present study might indicate the provision of high quality management by nephrologists, which might be partially supported by higher proportion of the patients receiving an ESA than those in the previous study .
Our risk score differs from previous scores with respect to target population and simplicity as well as availability of parameters in a routine nephrology care setting. For example, the prognostic score developed by Couchoud et al. was only applicable to elderly patients , whereas our score doesn’t include age. It may be counterintuitive that age was not associated with 1-year mortality. However, it is possible that effect of age on mortality was lessened by adjustment of performance status and comorbidity, as both are associated with aging. Regarding routine nephrology care settings, they included irrelevant patients with unplanned initiation of HD who have higher mortality than those with planned initiation of HD , and whether or not their patients received medical care by the nephrologists is unknown. In contrast, our risk score was developed from patients who received predialysis nephrology care and whose eGFR was then 10 mL/min per 1.73 m2 in an out-patient setting. Our score might therefore be more accurate than the previous scores for nephrologists when referring to personalized mortality risks to decide whether to initiate HD. The prognostic score developed by van Diepen et al. was also restricted to patients with diabetic nephropathy with a complicated score calculation requiring computer assistance . Chua et al. developed a 5-item prognostic tool with simple integer scores; however, this tool is difficult to use in routine settings due to the requirement of echocardiography measurement . In contrast to these previous studies, our risk score is applicable for CKD patients regardless of age and underlying disease and composed of simple integer scores available in daily nephrology practice settings.
The strength of this study is that the 6-item risk score for predicting 1-year mortality was developed using clinical data available immediately before the initiation of HD from multicenter nephrology institutions. As mortality tends to be relatively high one year after starting HD, this integer risk score might be useful for estimating prognosis, identifying high-risk patients, and making care plans. For example, when bedridden CKD patients with hemiplegia and peripheral vascular disease exhibit a low level of serum albumin and a high level of calcium, predicted 1-year mortality is >25% according to our risk score. The nephrologist would therefore be able to discuss HD initiation based on this predicted mortality with the patient and their family.
However, several limitations also warrant attention in the present study. First, CKD patients who started HD with an eGFR>10 mL/min per 1.73 m2 were excluded from this study. This might explain why emergent complications related to CKD, such as pulmonary edema, were not retained in the risk scores. Second, as our risk score was only validated internally, external validation should also be conducted. Third, 1-year mortality in the present population was lower than that reported in the Japanese registry (15.9%) . This lower mortality might be explained by the exclusion of patients with eGFR>10 mL/min per 1.73 m2 at initiation of HD, as a higher level of eGFR at the initiation of HD is associated with mortality [4, 18, 19, 32]. In addition, patients enrolled in the present study received predialysis nephrology care at least once, which has been related to improved survival after the initiation of HD [29, 33].
In conclusion, we developed and validated a simple risk score in CKD patients receiving predialysis nephrology care to predict 1-year mortality after the initiation of HD. This score might help nephrologists make personalized assessments for risk of mortality and facilitate making a shared decision with patients and families regarding the initiation of HD.
In addition to the authors, the following investigators participated in the study: I Nakaya and K Yoshikawa (Iwate Prefectural Central Hospital, Iwate, Japan), H Mukai and Y Matsubara (Department of Nephrology and Kidney Center, Aso-Iizuka Hospital, Fukuoka, Japan), T Ikenoue (Department of Healthcare Epidemiology, Graduate School of Medicine and Public Health, Kyoto University, Kyoto, Japan), K Koike (Division of Nephrology, Department of Medicine, Kurume University School of Medicine, Fukuoka, Japan), H Kawarazaki and N Tominaga (Division of Nephrology and Hypertension, St. Marianna University School of Medicine Hospital, Kanagawa, Japan), M Yazawa and T Tomino (Hakuyu Chiyoda Clinic, Osaka, Japan), N Okada (Chubu Rosai Hospital, Aichi, Japan), H Nishiwaki and M Takayasu (Division of Nephrology, Department of Medicine, Showa University Fujigaoka Hospital, Kanagawa, Japan), Y Tsuruta (Department of Medicine, Kidney Center, Tokyo Women's Medical University, Tokyo, Japan), H Morinaga and M Kitagawa (Department of Medicine and Clinical Science, Okayama University Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, Okayama, Japan), Y Wada (Division of Nephrology, Department of Medicine, Showa University School of Medicine, Tokyo, Japan), T Toyama (Obihiro Kyoukai Hospital, Hokkaido, Japan: K Nishino. Koryo Clinic, Toyama, Japan), T Suwabe and K Sumida (Nephrology Center, Toranomon Hospital, Tokyo, Japan), K Aoyagi (Hino Clinic, Osaka, Japan) and M Murakami (Shizuoka General Hospital, Shizuoka, Japan).
Conceived and designed the experiments: TD SY TM KS NK. Performed the experiments: TD SY TM KS NK. Analyzed the data: TD SY TM KS NK YO. Wrote the paper: TD SY TM KS NK YO.
- 1. Couchoud C, Labeeuw M, Moranne O, Allot V, Esnault V, Frimat L, et al. 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. pmid:19096087
- 2. Susantitaphong P, Altamimi S, Ashkar M, Balk EM, Stel VS, Wright S, et al. GFR at initiation of dialysis and mortality in CKD: a meta-analysis. Am J Kidney Dis. 2012;59: 829–840. pmid:22465328
- 3. Yamagata K, Nakai S, Masakane I, Hanafusa N, Iseki K, Tsubakihara Y, et al. Ideal timing and predialysis nephrology care duration for dialysis initiation: from analysis of Japanese dialysis initiation survey. Ther Apher Dial. 2012;16: 54–62. pmid:22248196
- 4. Wright S, Klausner D, Baird B, Williams ME, Steinman T, Tang H, et al. Timing of dialysis initiation and survival in ESRD. Clin J Am Soc Nephrol. 2010;5: 1828–1835. pmid:20634325
- 5. van Diepen M, Schroijen MA, Dekkers OM, Rotmans JI, Krediet RT, Boeschoten EW, et al. Predicting mortality in patients with diabetes starting dialysis. PLoS One. 2014;9: e89744. pmid:24594735
- 6. Matsuo S, Imai E, Horio M, Yasuda Y, Tomita K, Nitta K, et al. Revised equations for estimated GFR from serum creatinine in Japan. Am J Kidney Dis. 2009;53: 982–992 pmid:19339088
- 7. Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis. 1987;40: 373–383. pmid:3558716
- 8. Oken MM, Creech RH, Tormey DC, Horton J, Davis TE, McFadden ET, et al. Toxicity and response criteria of the Eastern Cooperative Oncology Group. Am J Clin Oncol. 1982;5: 649–655. pmid:7165009
- 9. van der Heijden GJ, Donders AR, Stijnen T, Moons KG. Imputation of missing values is superior to complete case analysis and the missing-indicator method in multivariable diagnostic research: a clinical example. J Clin Epidemiol. 2006;59: 1102–1109. pmid:16980151
- 10. van Buuren S, Boshuizen HC, Knook DL. Multiple imputation of missing blood pressure covariates in survival analysis. Stat Med. 1999;18: 681–694. pmid:10204197
- 11. Vergouwe Y, Royston P, Moons KG, Altman DG. Development and validation of a prediction model with missing predictor data: a practical approach. J Clin Epidemiol. 2010;63: 205–214. pmid:19596181
- 12. Sauerbrei W. The Use of Resampling Methods to Simplify Regression Models in Medical Statistics. Journal of the Royal Statistical Society: Series C (Applied Statistics) 1999;48: 313–329. pmid:12294883
- 13. Rubin DB. Multiple Imputation for Nonresponse in Surveys. New York: John Wiley & Sons; 1987. https://doi.org/10.1002/sim.6008 pmid:24122822
- 14. Moons KG, Donders AR, Steyerberg EW, Harrell FE. Penalized maximum likelihood estimation to directly adjust diagnostic and prognostic prediction models for overoptimism: a clinical example. J Clin Epidemiol. 2004;57:1262–1270 pmid:15617952
- 15. Harrell FE Jr, Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med. 1996;15: 361–387. pmid:8668867
- 16. Frank H. Using multiple imputation for Cox proportional hazards, then validating with rms package? Cross Validated. 20 Dec 2012. Available: http://stats.stackexchange.com/questions/46333/using-multiple-imputation-for-cox-proportional-hazards-then-validating-with-rms?rq=1. Accessed 01 Dec 2014.
- 17. Harrell FE. Regression Modeling Strategies: With Applications to Linear Models, Logistic Regression, and Survival Analysis. New York: Springer; 2001.
- 18. Kazmi WH, Gilbertson DT, Obrador GT, Guo H, Pereira BJ, Collins AJ, et al. Effect of comorbidity on the increased mortality associated with early initiation of dialysis. Am J Kidney Dis. 2005;46: 887–896. pmid:16253729
- 19. Clark WF, Na Y, Rosansky SJ, Sontrop JM, Macnab JJ, Glassock RJ, et al. Association between estimated glomerular filtration rate at initiation of dialysis and mortality. CMAJ. 2011;183: 47–53. pmid:21135082
- 20. Block GA, Klassen PS, Lazarus JM, Ofsthun N, Lowrie EG, Chertow GM. Mineral metabolism, mortality, and morbidity in maintenance hemodialysis. J Am Soc Nephrol. 2004;15: 2208–2218. pmid:15284307
- 21. Young EW, Albert JM, Satayathum S, Goodkin DA, Pisoni RL, Akiba T, et al. Predictors and consequences of altered mineral metabolism: the Dialysis Outcomes and Practice Patterns Study. Kidney Int. 2005;67: 1179–1187. pmid:15698460
- 22. Kalantar-Zadeh K, Kuwae N, Regidor DL, Kovesdy CP, Kilpatrick RD, Shinaberger CS, et al. Survival predictability of time-varying indicators of bone disease in maintenance hemodialysis patients. Kidney Int. 2006;70: 771–780. pmid:16820797
- 23. Floege J, Kim J, Ireland E, Chazot C, Drueke T, de Francisco A, et al. Serum iPTH, calcium and phosphate, and the risk of mortality in a European haemodialysis population. Nephrol Dial Transplant. 2011;26: 1948–1955. pmid:20466670
- 24. Palmer SC, Hayen A, Macaskill P, Pellegrini F, Craig JC, Elder GJ, et al. Serum levels of phosphorus, parathyroid hormone, and calcium and risks of death and cardiovascular disease in individuals with chronic kidney disease: a systematic review and meta-analysis. JAMA. 2011;305: 1119–1127. pmid:21406649
- 25. Verbeke F, Van Biesen W, Honkanen E, Wikström B, Jensen PB, Krzesinski JM, et al. Prognostic value of aortic stiffness and calcification for cardiovascular events and mortality in dialysis patients: outcome of the calcification outcome in renal disease (CORD) study. Clin J Am Soc Nephrol. 2011;6: 153–159. pmid:20829424
- 26. Russo D, Miranda I, Ruocco C, Battaglia Y, Buonanno E, Manzi S, et al. The progression of coronary artery calcification in predialysis patients on calcium carbonate or sevelamer. Kidney Int. 2007;72: 1255–1261. pmid:17805238
- 27. Fink J, Blahut S, Reddy M, Light P. Use of erythropoietin before the initiation of dialysis and its impact on mortality. Am J Kidney Dis. 2001;37: 348–355. pmid:11157377
- 28. Hayashi T, Kimura T, Yasuda K, Obi Y, Sasaki K, Iio K, et al. Prognostic significance of left ventricular hypertrophy observed at dialysis initiation depends on the pre-dialysis use of erythropoiesis-stimulating agents. Clin Exp Nephrol. 2013;17: 294–303. pmid:23100176
- 29. Bradbury BD, Fissell RB, Albert JM, Anthony MS, Critchlow CW, Pisoni RL, et al. Predictors of early mortality among incident US hemodialysis patients in the Dialysis Outcomes and Practice Patterns Study (DOPPS). Clin J Am Soc Nephrol. 2007;2: 89–99. pmid:17699392
- 30. Goodkin DA, Young EW, Kurokawa K, Prütz KG, Levin NW. Mortality among hemodialysis patients in Europe, Japan, and the United States: case-mix effects. Am J Kidney Dis. 2004;44: 16–21. pmid:15486869
- 31. Chua HR, Lau T, Luo N, Ma V, Teo BW, Haroon S, et al. Predicting first-year mortality in incident dialysis patients with end-stage renal disease—the UREA5 study. Blood Purif. 2014;37: 85–92. pmid:24589505
- 32. Beddhu S, Samore MH, Roberts MS, Stoddard GJ, Ramkumar N, Pappas LM, et al. Impact of timing of initiation of dialysis on mortality. J Am Soc Nephrol. 2003;14: 2305–2312. pmid:12937307
- 33. Hasegawa T, Bragg-Gresham JL, Yamazaki S, Fukuhara S, Akizawa T, Kleophas W, et al. Greater first-year survival on hemodialysis in facilities in which patients are provided earlier and more frequent pre-nephrology visits. Clin J Am Soc Nephrol. 2009;4: 595–602. pmid:19261827