E-Mail karger@karger.com

Original Paper

Nephron Clin Pract 2012;122:102–106 DOI: 10.1159/000350730

How Accurately Do Nephrologists Predict the Need for Dialysis within One Year?

Darren Green James P. Ritchie David I. New Philip A. Kalra

Vascular Research Group, Salford Royal Hospital, Manchester Academic Health Sciences Centre, Manchester , UK

sitive but not specific in predicting the need for dialysis. Ed-ucating the clinicians may improve the specificity of judge-ment and improve the accuracy of prognostic information given to patients. Copyright © 2013 S. Karger AG, Basel

Introduction

Predicting when patients with advanced chronic kid-ney disease (CKD) will require renal replacement therapy (RRT) is important. Accurate prediction can improve timing for vascular access and pre-emptive live-donor transplantation. Both of these can improve survival [1, 2] . Telling a patient they will require dialysis soon may influ-ence mood and lifestyle choices. It is therefore vital that predictions are accurate.

The current best practice guidelines on timing and ini-tiation of RRT acknowledge that the evidence supporting them is weak [3] . Previous cohort studies have reported which factors correlate with rapid progression of CKD. These have concentrated on biochemical results and medical history, and include earlier stages of CKD. Two studies have attempted to devise risk-assessment models using scoring based on such parameters. In the first, pa-tients in the highest risk quartile had a probability of pro-gression to RRT of 20% over 5 years [4] . In the second, development and validation cohorts produced a predic-

Key Words

Chronic kidney disease · Time to dialysis · Clinician judgement

Abstract

Background/Aims: Knowing when patients with chronic kidney disease will need dialysis can improve patient coun-selling and timing of vascular access. We aimed to assess the accuracy of clinician judgement in predicting the need for dialysis within 12 months. Methods: We asked the ne-phrologists in a dedicated pre-dialysis clinic to predict the time until initiation of dialysis for patients. We compared predicted with actual time to dialysis and the accuracy of predictions made by different grades of clinician. Multivari-ate logistic regression compared clinical parameters that correlated with predicted and actual time to dialysis. Re-

sults: One hundred and eighty-four patients were included. The sensitivity of clinician judgement as a predictor of di-alysis within 12 months was 95% and the specificity was 62%. Consultants were correct in 71% of cases and trainees in 68% of cases. Estimated glomerular filtration rate (eGFR) was the only independent correlate of predicted time to dialysis [odds ratio (OR) = 1.6 per 1 ml/min/1.73 m 2 reduc-tion, p < 0.001]. eGFR was also associated with actual time to dialysis (OR = 1.6 per 1 ml/min/1.73 m 2 , p < 0.001) along with age (OR = 0.94 per year increase, p = 0.005) and itch (OR = 3.7, p = 0.048). Conclusion: Clinical judgement is sen-

Received: April 26, 2012 Accepted: March 9, 2013 Published online: April 25, 2013

Darren Green Room C305, CSB, Salford Royal Hospital Manchester Academic Health Sciences Centre M6 8HD Manchester (UK) E-Mail darrengreen   @   doctors.org.uk

© 2013 S. Karger AG, Basel1660–2110/13/1224–0102$38.00/0

www.karger.com/nec

Accuracy in Predicting the Need for Dialysis

Nephron Clin Pract 2012;122:102–106DOI: 10.1159/000350730

103

tive model which used 13 measured parameters to give an individual probability of a patient ever reaching dialysis [5] . Other studies have shown that specific interventions slow CKD progression [6, 7] . However, these predomi-nantly focus on long-term progression, which is not use-ful in ensuring that preparations for RRT are made with-in an appropriate time frame. No study has prospectively evaluated progression in the context of pre-dialysis CKD, nor assessed clinician judgement as a predictive model rather than a combination of clinical parameters.

We hypothesised that clinician judgement is an effec-tive predictor of time to dialysis for individual patients. It is a patient-specific judgement that accommodates both the measured parameters implicated in previous studies as well as the subjective measure of symptom burden and other factors less easily quantified.

The secondary aim was to identify where weaknesses lay in clinicians’ ability to predict time to dialysis by com-paring the predictions of different grades of clinician, and whether the parameters that correlated with predictions were also those that correlated with actual time to dialysis. This, in turn, could allow for education, improved predic-tions and more timely interventions for patients.

Methods

A 3-month audit of all patients seen in a pre-dialysis clinic was performed. This clinic is designed to see patients with stage 5 CKD or glomerular filtration rate (GFR) ≤ 20 ml/min/1.72 m 2 and rap-idly progressive disease.

Nephrologists were asked to state whether they expected a pa-tient to be on dialysis within 12 months, and if so, to predict the time to dialysis in months. Judgement was made upon the clini-cal information available during the consultation. This included biochemical parameters and their historical trends, information contained in patient case-notes and the findings from the pa-tient’s history and examination. Patients who were referred to start dialysis upon completion of the consultation were excluded. Twelve months later, the case-notes were reviewed to determine which patients had reached dialysis. Patients who had died pre-dialysis, had been elected for conservative care, received a pre-emptive transplant or had had a live-donor graft scheduled at the time of consultation were excluded. Those who required haemo-filtration or early dialysis for acute kidney injury were also omit-ted as such intercurrent events cannot be predicted.

Sensitivity, specificity and positive and negative predictive val-ues (PPV and NPV) were calculated as a tool for clinician judge-ment to be able to predict dialysis within 1 year. These measure-ments were compared between grades of assessing clinician. In-dividual clinicians were not identifiable. A comparison was made between the assessments of consultants and of a group made up of nephrology registrars and a pre-dialysis specialist nurse.

Binary logistic regression was used to determine which of the clinical factors available to the clinician during the consultation

with a patient correlated with an expectation that that patient would be on dialysis within 1 year. If symptoms (oedema, fatigue and itch) were not recorded in the details of the consultation, they were presumed to be absent. Parameters significant at the level of α ≤ 0.05 on univariate analysis were included in a multivariate model to produce the final list of associated co-variates.

At 12 months, logistic regression was used to determine which clinical factors from the original consultation correlated with actual time to dialysis. The two sets of correlates were com-pared to display where clinicians were over or under-emphasis-ing the importance of particular clinical factors in predicting time to dialysis.

Results

Cosultation took place with 216 patients, 32 of whom were excluded (24 deaths, 2 clinic admissions, 3 cases of acute kidney injury and 3 transplants). This left 184 in the study. The mean age was 63.6 ± 15.1 years, mean estima-ted (e)GFR 13.8 ± 4.0 ml/min/1.73 m 2 , 65% of the patients were male and 41 (22%) had started dialysis by 12 months. There was no difference in any clinical parameter be-tween the patients seen by consultants and those seen by trainees.

Ninety-three patients (51%) were predicted to be on dialysis within 12 months, 39 (42%) of whom were. The estimated time to dialysis was correct to within 6 months in 90% of the cases. Of 91 patients who were expected to still be pre-dialysis at 12 months, only 2 (2%) had started dialysis.

The sensitivity of clinician judgement in predicting di-alysis within 12 months was 95% and the specificity was 62%. NPV was 98% but PPV was 42%. Consultants’ pre-dictions were correct in 71% of the cases and non-consul-tant grades were correct in 68%. There was a difference in PPV: for consultants this was 51% compared with 35% for non-consultants ( table  1 ). The accuracy of predictions also varied with the age of the patient. In the group of pa-

Table 1. Sensitivity, specificity, PPV and NPV of clinician judge-ment in predicting dialysis within 12 months

Grade of clinicianAll Consultant Trainee

Sensitivity, % 95 95 95Specificity, % 62 61 63PPV, % 42 51 35NPV, % 98 97 98

Green/Ritchie/New/Kalra Nephron Clin Pract 2012;122:102–106DOI: 10.1159/000350730

104

tients <70 years of age, of all those expected to be on di-alysis within 12 months, 57% were. This compared with 38% of patients aged ≥ 70 years.

Univariate analysis showed that eGFR, serum albumin, serum phosphate and urine protein:creatinine ratio all had statistically significant links with the predicted need for dialysis. Only eGFR showed significance on multivar-iate regression [odds ratio (OR) = 1.6 per 1 ml/min/1.73 m 2 reduction, p < 0.001].

For actual need for dialysis, multivariate analysis showed that eGFR again showed correlation (OR = 1.6 per 1 ml/min/1.73 m 2 , p < 0.001). In addition, age (OR = 0.94 per year increase in age, p = 0.005) and uraemic itch (OR = 3.7, p = 0.048) were also associated with dialysis within 12 months ( table  2 ). Importantly, it was eGFR rather than historical change in eGFR in the preceding 12 months before the judgement was made (δeGFR) that was different between the patients who did and did not need dialysis within a year. Baseline eGFR was 10.6 ± 2.9 and 15.0 ± 3.7 ml/min/1.73 m 2 , respectively. For δeGFR, the figures were –0.41 ± 1.54 and –0.49 ± 1.97 ml/min/1.73 m 2 /year, respectively.

When only assessing the characteristics of patients who were predicted to be on dialysis ( table 3 ), patients who did require dialysis were, again, younger and had a lower eGFR at consultation. There were numerical differ-ences in proteinuria and serum phosphate that did not reach statistical significance.

An assessment was made of the number of patients who died pre-dialysis with dialysis access in place: this was the case in 13 out of the 24 (54%) deaths during fol-low-up. This compares with 33 of 143 (23%) who either did not need dialysis or died. Pre-dialysis death with ac-

cess was largely a function of patients who had chosen haemodialysis as a modality. Here, 83 patients had either died during the study period or were still pre-dialysis after 12 months; an arteriovenous fistula had formed in 53 (64%) of these. There were 16 (22%) deaths in the case of patients who had chosen haemodialysis but for whom it had not yet been initiated; 13 (81%) of these had fistulae.

Table 2. Multivariate logistic regression analysis of baseline co-variates and their relationship with predicted and actual time to dialysis

Predicted ActualOR sig OR sig

Age, years – – 0.94 (0.90–0.98) 0.005eGFR, ml/min/1.73 m2 0.62 (0.49–0.78) 0.001 0.64 (0.50–0.83) 0.001Urine PCR, mg/mmol 1.01 (0.99–1.02) 0.10 1.01 (0.99–1.01) 0.59Serum albumin, g/dl 0.99 (0.84–1.21) 0.94 – –Serum phosphorus, mg/dl 2.53 (0.29–21.8) 0.39 1.46 (0.21–10.2) 0.70Presence of itch – – 3.74 (1.61–10.4) 0.04

Variables included in analyses are those that were significant at the α ≤ 0.05 level. OR (odds ratio) is per unit increase in variable or for the presence of categorical variables. Figures in parentheses represent 95% confidence intervals. PCR = Protein:creatinine ratio; sig = significance.

Table 3. Baseline demographics of patients predicted to be on dialysis within 12 months

Total Status at 12 monthsPre- dialysis Dialysis

Number of patients 93 54 39Mean age, years 61.5±14.7 64.9±13.3 58.4±15.1Male, % 68 68 68Diabetes, % 30 32 29Hypertension, % 83 91 75CAD, % 24 27 21eGFR, ml/min/1.73 m2 11.0±2.7 12.1±2.7 9.9±2.1δeGFR, ml/min/1.73 m2 –0.25±1.40 –0.24±1.57 –0.27±1.47Haemoglobin, g/dl 11.4±1.5 11.4±1.2 11.3±1.6PTH, pg/ml 212±151 202±185 221±110Albumin, g/dl 4.2±0.5 4.3±0.5 4.2±0.4Phosphorus, mg/dl 4.6±0.9 4.4±0.9 4.9±0.9Urine PCR, mg/mmol 230±196 184±155 272±225

This compares the features of those in whom the prediction was correct versus those in whom it was not correct. Values are displayed as a percentage or as mean values. Mean values are dis-played with 95% confidence intervals. CAD = Coronary artery disease; δeGFR = change in eGFR in preceding 12 months; PCR = protein:creatinine ratio; PTH = parathyroid hormone.

Accuracy in Predicting the Need for Dialysis

Nephron Clin Pract 2012;122:102–106DOI: 10.1159/000350730

105

Patients who died with fistulae present were older (71.1 ± 7.6 vs. 63.7 ± 6.9 years, p = 0.027) but with a higher eGFR (12.9 ± 1.5 vs. 11.1 ± 2.0 ml/min/1.73 m 2 ) than those pa-tients with fistulae who were still in the pre-dialysis clinic at 12 months. None of the patients who had chosen peri-toneal dialysis died with a catheter in place.

Discussion

Unlike previous studies, we specifically assessed time to dialysis in a cohort of pre-dialysis patients. We assessed clinician judgement as a tool to predict progression to di-alysis, rather than a model which depended on measured parameters alone. Clinician judgement is sensitive for predicting dialysis within 12 months (95%), but is only moderately specific (62%). Only 6% of patients in this co-hort reached dialysis 6 months sooner than expected. These results suggest clinician judgement is a generally safe clinical tool in avoiding unexpected initiation of di-alysis. A high rate of correct predictions should be ex-pected from clinicians as they will ultimately go on to make decisions to start dialysis. This confirmation bias can be partially refuted by the fact that the non-consultant grades, which are unable to list patients to commence di-alysis at our centre, had an equal sensitivity to the consul-

tants. The more pertinent message is that clinicians over-predict the likelihood of patients needing dialysis within the next 12 months. False positives accounted for 29% of all predictions, and 58% of those expected to be on dialy-sis within a year were not.

The finding that patients from a pre-dialysis clinic who have chosen haemodialysis as an RRT modality were more likely to die with a fistula than without goes some way to quantify the dilemma of early vascular access. This does not add weight to the argument against early access but highlights the need for frank counselling on this sub-ject. This is particularly the case for elderly patients in whom the rate of progression of CKD may be slower [6, 8] and the mortality rate higher.

In this study, patients ≥ 70 years old who were expect-ed to reach dialysis within a year were almost as likely to die (22 vs. 30%), and false positive predictions were high-er than in the <70 years age group. Our results suggest that it may not be appropriate to have an early discussion regarding RRT and arrange early vascular access for el-derly patients. What is clear is that consultants make few-er false positive predictions. Perhaps dialysis and vascular access planning should only be made by senior clinicians.

Within this cohort, nephrologists correctly identified eGFR as a strong predictor of needing dialysis. However, they did not appear to make adequate use of the associa-

Table 4. Predictors of progression of CKD in previous cohort studies

Landray [9] Nakamura [6] Agarwal [10] Levin [11] Conway [8] Hoefield [12] Our study

CKD stagesPatients 3–5 any 3–5 4–5 4 3–5 eGFR <20

Number 382 366 220 4,231 396 1,325 184Age no older no younger younger younger youngerGender female n.a. n.a. male no no noHypertension no yes yes yes n.a. yes noSmoking no n.a. n.a. n.a. n.a. no n.a.CAD no no no n.a. n.a. no noDiabetes no no n.a. no no no noeGFR yes yes yes yes yes yes yesδeGFR no n.a. yes n.a. yes no noPhosphorus yes yes n.a. yes n.a. yes noAnaemia no yes n.a. yes yes yes noProteinuria yes n.a. yes yes yes yes noSymptoms n.a. n.a. n.a. n.a. n.a. n.a. itchAlbumin no n.a. n.a. n.a. n.a. no no

CAD = Coronary artery disease; δeGFR = time-dependent change in eGFR; n.a. = variables that were not assessed.

Green/Ritchie/New/Kalra Nephron Clin Pract 2012;122:102–106DOI: 10.1159/000350730

106

tion of other parameters, namely age and itch. This fact indicates that there is scope for improvement in predic-tions, and ongoing education may improve rates of suc-cessful prediction. A potential flaw is that with reduced numbers of patients being expected to reach dialysis, there may be an increase in the number of patients reach-ing dialysis sooner than expected (a higher rate of false negative predictions).

In our cohort, fewer parameters were predictive of time to dialysis than in previous studies ( table 4 ). This may be an effect of sample size, our biggest shortcoming. Whilst there are some differences across studies, they largely agree that younger patients with a lower eGFR require

RRT the soonest. Other studies show that other laboratory results (e.g. phosphate, haemoglobin and proteinuria) are also predictive. That these did not appear in our multi-variate model may be due to sample size, reflect the differ-ent populations between studies or reflect our inclusion of symptoms that may relate to the laboratory findings.

In this study, a symptom of advanced CKD, namely itch, was predictive of the early need for dialysis. This has not been expressed in previous studies and indicates that symptom burden should be accommodated when decid-ing on the timing of access to dialysis and should be in-cluded in future studies that attempt to produce a predic-tive model.

References

1 Perl J, Wald R, McFarlane P, Bargman JM, Vonesh E, Na Y, Jassal SV, Moist L: Hemodi-alysis vascular access modifies the association between dialysis modality and survival. JASN 2011; 22: 1113–1121.

2 Vollmer WM, Wahl PW, Blagg CR: Survival with dialysis and transplantation in patients with end-stage renal disease. N Engl J Med 1983; 308: 1553–1558.

3 Dombros N, Dratwa M, Feriani M, Gokal R, Heimbürger O, Krediet R, Plum J, Rodrigues A, Selgas R, Struijk D, Verger C, EBPG Expert Group on Peritoneal Dialysis: European Best Practice Guidelines: 2. The initiation of dialy-sis. Nephrol Dial Transplant 2005; 20: 3–7.

4 Johnson ES, Thorp ML, Platt RW, Smith DH: Predicting the risk of dialysis and transplant among patients with CKD: a retrospective co-hort study. Am J Kidney Dis 2008; 52: 653–660.

5 Tangri N, Stevens LA, Griffith J, Tighiouart H, Djurdjev O, Naimark D, Levin A, Levey AS: A predictive model for progression of chronic kidney disease to kidney failure. JAMA 2011; 305: 1553–1559.

6 Nakamura S, Nakata H, Yoshihara F, Kamide K, Horio T, Nakahama H, Kawano Y: Effect of early nephrology referral on the initiation of hemodialysis and survival in patients with chronic kidney disease and cardiovascular diseases. Circ J 2007; 71: 511–516.

7 Peterson JC, Adler S, Burkart JM, Greene T, Hebert LA, Hunsicker LG, King AJ, Klahr S, Massry SG, Seifter JL: Blood pressure control, proteinuria, and the progression of renal dis-ease. The Modification of Diet in Renal Dis-ease Study. Ann Intern Med 1995; 123: 754–762.

8 Conway B, Webster A, Ramsay G, Morgan N, Neary J, Whitworth C, Harty J: Predicting mortality and uptake of renal replacement therapy in patients with stage 4 chronic kid-ney disease. Nephrol Dial Transplant 2009; 24: 1930–1937.

9 Landray MJ, Emberson JR, Blackwell L, Das-gupta T, Zakeri R, Morgan MD, Ferro CJ, Vickery S, Ayrton P, Nair D, Dalton RN, Lamb EJ, Baigent C, Townend JN, Wheeler DC: Prediction of ESRD and death among people with CKD: the Chronic Renal Impair-ment in Birmingham (CRIB) prospective co-hort study. Am J Kidney Dis 2010; 56: 1082–1094.

10 Agarwal R, Bunaye Z, Bekele DM, Light RP: Competing risk factor analysis of end-stage renal disease and mortality in chronic kidney disease. Am J Nephrol 2008; 28: 569–575.

11 Levin A, Singer J, Thompson C, Ross H, Lew-is M: Prevalent left ventricular hypertrophy in the predialysis population: identifying op-portunities for intervention. Am J Kidney Dis 1996; 27: 347–354.

12 Hoefield R, Kalra PA, Baker P, Lane B, New J, O’Donoghue D, Foley RN, Middleton R: Fac-tors associated with kidney disease progres-sion and mortality in a referred CKD popula-tion. Am J Kidney Dis 2010; 56: 1072–1081.