- Email: [email protected]
Simple nutritional indicators as independent predictors of mortality in hemodialysis patients
Simple Nutritional Indicators as Independent Predictors of Mortality in Hemodialysis Patients Sean F. Leavey, MB, MRCP, Robert L. Strawderman, ScD, Camille A. Jones, MD, MPH, Friedrich K. Port, MD, MS, and Philip J. Held, PhD ● A strong association exists between nutritional status and morbidity and mortality in patients with end-stage renal disease who are treated with hemodialysis. Described here is the predictive value for mortality over 5 years of follow-up of a number of risk factors, recorded at baseline, in a national sample of 3,607 hemodialysis patients. Among the variables studied were case-mix covariates, caregiver classifications of nutritional status, serum albumin concentration, and body mass index (BMI). The Case Mix Adequacy special study of the United States Renal Data System (USRDS) provided these measurements as of December 31, 1990. The USRDS patient standard analysis file provided follow-up data on mortality for all patients through December 31, 1995, by which time 64.7% of the patients had died. BMI is a simple anthropometric measurement that has received little attention in dialysis practice. Caregiver classification refers to documentation in a patient’s dialysis facility medical records that stated an individual to be ‘‘undernourished/cachectic,’’ ‘‘obese/overweight,’’ or ‘‘well-nourished.’’ The mean serum albumin was 3.7 ⴞ 0.45 (SD) g/dL, and the mean BMI was 24.4 ⴞ 5.3 (SD) kg/m2. By caregiver classification, 20.1% of patients were undernourished, and 24.9% obese. In hazard regression models, including but not limited to the Cox proportional hazards model, low BMI, low serum albumin, and the caregiver classification ‘‘undernourished’’ were independently and significantly predictive of increased mortality. In analyses allowing for time-varying relative mortality risks (ie, nonproportional hazards), the greatest predictive value of all three variables occurred early during follow-up, but the independent predictive value of baseline serum albumin and BMI measurements on mortality risk persisted even 5 years later. No evidence of increasing mortality risk was found for higher values of BMI. Serum albumin was confirmed to be a useful predictor of mortality risk in hemodialysis patients; BMI was established as an independently important predictor of mortality; both serum albumin and BMI, measured at baseline, continue to possess predictive value 5 years later; the subjective caregiver classification of nutritional status ‘‘undernourished’’ has independent value in predicting mortality risk beyond the information gained from two other markers of nutritional status—BMI and serum albumin. r 1998 by the National Kidney Foundation, Inc. INDEX WORDS: Nutritional status; body mass index; serum albumin; hemodialysis; end-stage renal disease; mortality; hazard regression.
A
HIGH INCIDENCE of mortality complicates the treatment of hemodialysis patients with end-stage renal disease (ESRD). Protein and caloric intakes in maintenance hemodialysis patients have been observed to fall, on average, below the recommended requirements,1,2 and protein-energy malnutrition is considered to be a major determinant of morbidity and mortality.3-14 Several studies have documented a strong relationship between serum albumin concentration and mortality.7,8,10-14 These observations have been in both prevalent and incident ESRD patients. Serum albumin is considered to reflect visceral protein stores and thus to act as a marker of protein malnutrition. Energy malnutrition is more difficult to estimate, and whether it independently predicts an increased mortality risk in maintenance hemodialysis patients is unclear. The anthropometric measurement body mass index (BMI), calculated as an individual’s weight in kilograms divided by the square of their height in meters (kg/m2), correlates sufficiently with direct measures of body
fatness15 (eg, as measured by hydrodensitometry), to be used as a marker of energy nutritional status (energy storage). Data collected by the United States Renal Data System (USRDS) Case Mix Adequacy study, on a randomly selected national sample of Medicare ESRD patients who were treated with maintenance hemodialysis, provided a unique opportunity to examine the distribution and independent predictive value for mortality, controlling for case mix (eg, age, race, gender, and various comorbidities) of baseline measures of BMI, serum albumin concentration,
From the United States Renal Data System and the Kidney Epidemiology and Cost Center, Departments of Internal Medicine and Biostatistics, University of Michigan, Ann Arbor, MI and Bethesda, MD. Received September 30, 1997; accepted in revised form December 19, 1997. Address reprint requests to Sean F. Leavey, MB, MRCP, Kidney Epidemiology and Cost Center, 315 W Huron St, Suite 240, Ann Arbor, MI 48103. E-mail: [email protected]
r 1998 by the National Kidney Foundation, Inc. 0272-6386/98/3106-0014$3.00/0
American Journal of Kidney Diseases, Vol 31, No 6 (June), 1998: pp 997-1006
997
998
LEAVEY ET AL
and a third nutritional indicator, the nutritional status as documented in the medical records, by the clinician/caregiver. All three nutrition-related parameters, measured at baseline, were shown in this study to independently predict mortality risk. The effects of baseline BMI and serum albumin remained important up to 5 years after their date of measurement.
Table 1. Demographic Characteristics and Distributions of Baseline Comorbid Covariate Values for the Overall Random Sample Population (n ⴝ 6,585) and the Survival Analysis Subpopulation (n ⴝ 3,607)
Variable*
MATERIALS AND METHODS
Data The data used in this analysis were collected for the USRDS Case Mix Adequacy study and were supplemented with additional mortality follow-up data available from the USRDS patient database. The Case Mix Adequacy study collected data on a random sample of Medicare patients who were alive on hemodialysis on the study start date (December 31, 1990). The ‘‘duration of ESRD prior to study start date’’ varied significantly within the study population and was accounted for in all analyses. The data were abstracted retrospectively from the patients’ dialysis facility medical records by the 18 ESRD Networks under a contract with the Health Care Financing Administration (HCFA) of the U.S. Department of Health and Human Services. A copy of the data abstraction form for this study is published in the USRDS 1996 Annual Data Report, Appendix B. Details abstracted from each patient’s dialysis facility medical record included patient demographics; the presence or absence of a variety of comorbid health conditions occurring within 10 years before the study start date (Table 1); patient height (recorded any time); patient dry weight; serum albumin concentration; and documented nutritional status. The dry weight and documented caregiver classification of nutritional status were obtained from ⫾1 month of the study start date (December 1990 and January 1991). If dry weight was not clearly documented within that period, then the lowest weight within ⫾2 weeks of the study start date was recorded as the dry weight. BMI was computed from an individual’s dry weight in kilograms divided by the square of their height in meters (kg/m2). The serum albumin concentration was taken as the average of at least two values obtained within ⫾1 month of the study start date. Supplemental information on mortality through December 31, 1995, were obtained from the main USRDS patient database. To compare the relative size (as measured through BMI) of ESRD patients with their healthier counterparts in the general US population, we obtained data on BMI for the latter from the National Health and Nutrition Examination Survey III (NHANES III) phase 1 study.16 NHANES III, the third in a series of cross-sectional surveys conducted by the National Center for Health Statistics, was conducted between 1988 and 1994. It examined a nationally representative sample of the US civilian noninstitutionalized population and is an excellent source of information on a variety of health-related measures, including BMI.
Demographics Mean age, years (SD) Race (% black) Gender (% female) Vintage, years (SD)† Nutritional parameters Mean BMI, kg/m2 (SD) Mean serum albumin, g/dL (SD) ‘‘Undernourished’’ (caregiver classification‡) % ‘‘Obese’’ (caregiver classification‡) % Comorbid conditions (% yes) Diabetes Coronary artery disease§ Congestive heart failure Left ventricular hypertrophy Peripheral vascular disease\ Smoking Pulmonary disease¶ Functional status (% yes) Unable to ambulate independently
Full Database (Max n ⫽ 6,585) Mean (SD) or Percent
Survival Analysis (n ⫽ 3,607) Mean (SD) or Percent
58.5 (16) 35.9 49.6 3.4 (3.8)
57.1 (16) 37.1 50.0 3.4 (3.8)
24.3 (5.8)
24.4 (5.3)
3.7 (0.5)
3.7 (0.5)
22.2
20.1
24.3
25.0
40.0 46.0 43.2
41.2 47.7 43.7
39.0
40.9
22.2 16.8 11.4
23.0 17.1 11.6
8.3
7.6
*All of these variables are included in the survival analysis. †Vintage refers to the mean duration of ESRD before study start date. ‡Includes documentation in the medical records of ‘‘undernourished/cachectic,’’ ‘‘obese/overweight,’’ or ‘‘well nourished’’ within 1 month of study start date. §Includes history of angina, myocardial infarction, coronary angioplasty, coronary artery bypass surgery, or abnormal coronary angiography. \Includes history of claudication, amputation, absent foot pulses, and prior diagnosis of peripheral vascular disease. ¶Includes chronic obstructive pulmonary disease, asthma, and use of home oxygen.
Analyses Baseline distributions for the three major nutritional parameters and several important indicators of case-mix severity were determined for the total possible sample population in the Case Mix Adequacy study (n ⫽ 6,585). Comparison
NUTRITION AND MORTALITY IN HEMODIALYSIS
was made with their respective distributions for the 3,607 hemodialysis patients who were included in the current mortality analyses, and who had no missing values for any of the nutritional parameters. A one-sided, two-sample t-test was used to compare age-specific mean BMI for men and women in the Case Mix Adequacy population to the mean BMI for men and women in the general US population 20 years of age or older. Missing values for the three major nutritional parameters primarily accounted for the decreased sample size used in the mortality analyses as follows: 1,140 patients were missing BMI (mostly because of missing height); 1,696 patients were missing documentation of their nutritional status within 1 month of December 31, 1990; and 293 patients were missing documentation of serum albumin levels. In addition, 636 patients with neoplasm and 116 patients with cirrhosis were also excluded from the analysis, because it was thought that including both of these factors might adversely impact a fair evaluation of the nutrition-related measures on mortality. Exclusion for one or more of these reasons left a final sample size of 3,607 patients. The distribution of BMI in our sample was found to be highly skewed, with a long, right-hand tail. Hazard regression models, including but not limited to Cox proportional hazards regression, were used to evaluate the relationship between BMI and mortality in this study.17-19 It can be shown mathematically that the influence of any single observation on the estimated regression coefficient in most regression models depends on the distance of that individual’s covariate value from the sample mean value; such is the case, for example, in Cox proportional hazards regression.20 Consequently, to avoid a situation in which a few high BMI observations could easily have had an undue influence on the results, we used a natural log transformation of BMI (lnBMI); the latter was essentially normally distributed about its mean value. In the mortality analyses, all continuous variables (save ‘‘duration of ESRD prior to study start date’’) were centered at their mean values; hence, we measured unit changes in lnBMI, serum albumin, and age (1.0 units, 1 g/dL, 1 year, respectively) relative to their respective mean values (ie, relative risk [RR] ⫽ 1.0 at the mean). Exploratory mortality analyses, performed for the purposes of model building and proportional hazard assessments, were done using both HARE (a nonparametric hazard regression program) and time-dependent coefficient hazard regression models.18,19 Using HARE, no significant evidence of nonlinear effects on the RR for mortality was found for any of the continuous variables used in the mortality analysis, save ‘‘duration of ESRD prior to study start date.’’ The increase in the RR for mortality associated with each year increase in the ‘‘duration of ESRD prior to study start date’’ slowed substantially after 2 years, the baseline group having a ‘‘duration of ESRD prior to study start date’’ of 0 years. A linear spline function for ‘‘duration of ESRD prior to study start date,’’ with a single breakpoint at 2 years, was used in the final reported Cox regression model to account for this change in RR. Significant interactions between age and diabetes, and black race and diabetes, were identified and are included in the final model. In addition to the HARE models, the use of timedependent coefficient models19 strongly indicated nonpropor-
999
tional hazards (non-PH) effects over time of congestive heart failure (CHF), serum albumin, ‘‘undernourished,’’ and lnBMI. Stratifying on CHF conveniently addressed the nonproportional effects of this variable without resulting in unbalanced strata. This was not an option for the latter three variables because their effects were of direct interest in this investigation. A robust variance estimate was used to help account for the fact that there were known imperfections (eg, due to non-PH) in the final reported Cox proportional hazards model.21 Plots were generated to show (1) the changes in the estimated relative mortality risk over time for serum albumin, lnBMI, and ‘‘undernourished’’ obtained under the time-dependent coefficient model19; and (2) the relationship between BMI and mortality risk (through an estimated RR for lnBMI) suggested by a Cox proportional hazards model. A sensitivity analysis, to verify the linear effect of lnBMI on lnRR, was done where BMI was categorized into quintiles, such that there were an equal number of deaths in each quintile. A Cox model was fit, replacing lnBMI with indicators for each of the lower and higher two quintiles (the middle quintile constituting the reference group, that is, those assigned an RR of 1.0). Supplemental analyses were done to evaluate the effects of the explanatory variables separately in those patients with and without CHF. Supplemental models also looked at the effects of the caregiver classification ‘‘obese.’’ An independent effect of obese and BMI was believed to be implausible, and to investigate this, we evaluated the effects of obesity on mortality risk with and without adjusting for BMI. All mortality analyses were done using the S-Plus Statistical Software System version 3.3, (StatSci, Seattle, WA). For all analyses, statistical significance is reported at the ␣ ⫽ 0.05 level.
RESULTS
The demographic characteristics and baseline comorbid covariates for both the total random sample Case Mix Adequacy study population (n ⫽ 6,585) and the subpopulation included in the mortality analysis (n ⫽ 3,607) are shown in Table 1. For all variables, the subpopulation was similar to the total population. Comparisons of BMI among male and female hemodialysis patients and the general US population, by age group, are shown in Table 2. In all age groups and both genders mean BMI was significantly lower in hemodialysis patients than in the respective general population samples. The distribution of BMI in the sample was found to be highly skewed, with a long righthand tail. The distributions of serum albumin and lnBMI, however, were approximately normally distributed. Serum albumin and lnBMI also showed a very weak correlation of r ⫽ 0.04. These results suggest that lnBMI and serum albumin essentially varied independently of one
1000
LEAVEY ET AL
Table 2. Comparison of Age-Specific Mean Body Mass Index for Hemodialysis Patients in Case Mix Adequacy Study Versus General US Population in NHANES III Phase I Study by Gender General Population*
Age (yr)
Men 20-29 30-39 40-49 50-59 60-69 70-79 ⬎80 Women 20-29 30-39 40-49 50-59 60-69 70-79 ⬎80
Hemodialysis Population
Sample Size
Mean BMI (SE)
Sample Size
Mean BMI (SE)
P
858 759 643 493 588 495 373
24.9 (0.21) 26.1 (0.29) 27.3 (0.36) 27.6 (0.16) 26.9 (0.22) 26.5 (0.29) 24.7 (0.24)
134 327 377 490 675 523 141
22.1 (0.29) 23.5 (0.26) 25.0 (0.28) 24.8 (0.23) 23.9 (0.16) 22.9 (0.16) 21.7 (0.29)
⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001
755 771 624 464 595 446 396
24.1 (0.29) 26.4 (0.39) 26.7 (0.29) 28.5 (0.41) 27.3 (0.27) 26.7 (0.29) 24.6 (0.23)
134 201 312 448 804 613 164
22.7 (0.51) 24.1 (0.42) 25.8 (0.36) 26.1 (0.27) 25.3 (0.20) 23.4 (0.20) 21.9 (0.29)
0.0103 ⬍0.0001 0.0284 ⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001
*Reference 16.
another, and consequently, any statistically significant effects of lnBMI or serum albumin on mortality were unlikely to be due to the other being significant. All of the reported data on the relative risks for mortality associated with changes in any one of the three nutritional parameters, BMI, serum albumin, and ‘‘undernourished,’’ adjusted both for the effects of the other two nutritional parameters and the following case mix variables: age; race; gender; ‘‘duration of ESRD prior to study start date’’; diabetes; coronary artery disease; left ventricular hypertrophy; peripheral vascular disease; smoking status; pulmonary disease; and ambulatory status. Unless otherwise specified, the models were stratified on CHF. In all, 64.7% of patients died over the 5-year follow-up period. The results of a Cox proportional hazards analysis of mortality risk are shown in Table 3. A lower mortality risk accompanied higher values of both lnBMI and serum albumin, relative to their mean values. An increased RR for mortality accompanied the caregiver classification ‘‘undernourished.’’ The effect of each nutritional parameter on mortality risk was seen to
be highly statistically significant (P ⬍ 0.0005 in each instance) even when adjusting for the other two. The effects of each of these parameters were also found to be time dependent. The timedependent changes in RR are plotted in Figure 1A through C. For each variable, both the fixed RR from the Cox proportional hazards model of Table 3 and the time-varying RR with 95% confidence intervals estimated from the timedependent coefficient model are shown. The effect on mortality risk associated with baseline measurements of ‘‘undernourished,’’ serum albumin, and lnBMI was greatest at the beginning, approached the estimated RR obtained under the Cox proportional hazards model during the second year, and, for serum albumin and lnBMI, remained statistically significant up to 5 years after the measurement was taken. A baseline diagnosis of ‘‘undernourished’’ increased the risk of mortality in the first 3 years; thereafter, the confidence intervals cross the RR of 1.0 (Fig 1C). Figure 1A through 1C illustrates that the RRs reported in the Cox proportional hazards model for serum albumin, BMI, and ‘‘undernourished’’ (Table 3) might be best regarded as ‘‘timeweighted’’ averages. Figure 2 contains a plot of the RR for mortality as a function of BMI. The solid line in this plot shows a decreasing relative risk of mortality with an increase in BMI over the range of values seen in this study, or 15 to 45 kg/m2. All other variables being equal, this range of BMI measurements corresponds to a range in the estimated RR for mortality of 1.2 to 0.8 (RR of 1.0 being centered at the geometric mean BMI for the study population, 23.9 kg/m2). The solid line was generated from the fitted Cox model described in Table 3, which treats BMI as a continuous variable. The model fit assumes that the RR is linear (on the natural log scale) in lnBMI. Superimposed on this plot are the results of a separate sensitivity analysis. The intent of this sensitivity analysis was to evaluate whether the solid line, which is based on the above parametric assumption, reasonably represents the relationship between mortality risk and BMI. Five categories of BMI were used in this sensitivity analysis and corresponded to the quintiles of BMI among those who died. BMI was entered as a set of four categorical variables, rather than one continuous variable, into an otherwise unchanged Cox pro-
NUTRITION AND MORTALITY IN HEMODIALYSIS
1001
Table 3. Proportional Hazards Relative Mortality Estimates for Hemodialysis Patient Risk Factors Present at Study Start Date (Stratified for Congestive Heart Failure)
Demographic, Nutritional, or Comorbid Factors and Interactions
Regression Coefficient
Robust SE
Relative Mortality Rate
P
Age (per 1 year older) Race (black v nonblack) Gender (female v male) Vintage 0-2 yr (per yr) Vintage ⬎2 yr (per yr) LnBMI (per 1 unit, eg, 3.7 v 2.7)* Serum albumin (per 1.0 g/dL higher) Undernourished (yes v no) Diabetic (yes v no) Coronary artery disease (yes v no) Left ventricular hypertrophy (yes v no) Peripheral vascular disease (yes v no) Smoking, history of (yes v no) Nonambulatory (yes v no) Pulmonary disease (yes v no) Interactions: Age ⴱ diabetes (v age ⴱ nondiabetic) Black ⴱ diabetes (v black ⴱ nondiabetic)
0.0374 ⫺0.0567 ⫺0.0375 0.1272 ⫺0.1142 ⫺0.4470 ⫺0.3950 0.2223 0.5563 0.2274 0.0896 0.2959 0.1776 0.4436 0.2373
0.0022 0.0671 0.0449 0.0282 0.0318 0.1207 0.0530 0.0610 0.0636 0.0473 0.0449 0.0542 0.0637 0.0815 0.0706
1.04 0.95 0.96 1.14 0.89† 0.64 0.67 1.25 1.74 1.26 1.09 1.34 1.19 1.56 1.27
⬍0.0001 0.40 0.40 ⬍0.0001 ⬍0.0001 0.0002 ⬍0.0001 0.0003 ⬍0.0001 ⬍0.0001 0.046 ⬍0.0001 0.0053 ⬍0.0001 0.0008
⫺0.0097 ⫺0.2247
0.0033 0.0913
0.99 0.80
0.0029 0.014
*BMI 40.4 v 14.9, see Figure 2 for BMI. †RR ⫽ 1.01 per year after 2 years.
portional hazards model (see Table 3). The reference group was taken to be the middle quintile, which by definition had an RR of 1.0. A significantly higher risk for mortality was seen in the lowest quintile group (P ⫽ 0.02), and a marginally insignificant lower risk for mortality (P ⫽ 0.098) was seen in the highest quintile group. The fact that the solid line lies well within the 95% confidence limits obtained from the model treating BMI as a set of four categorical variables indicates that the solid line does indeed provide a concise and reasonable description of the relationship between mortality risk and BMI. Two significant interactions were detected: the increasing RR of mortality with increasing age, although still present, was less pronounced in diabetics than in nondiabetics; and the protective effect of being black was greater in diabetics than in nondiabetics. The main model adjusted for both of these interactions (Table 3). A spline function describes the effect of ‘‘duration of ESRD prior to study start date.’’ The increasing RR for mortality is steeper for ‘‘duration of ESRD prior to study start date’’ up to 2 years and less steep for ‘‘duration of ESRD prior to study start date’’ greater than 2 years. The main model therefore allows for differential effects on mortal-
ity for ‘‘duration of ESRD prior to study start date’’ up to 2 years and ‘‘duration of ESRD prior to study start date’’ greater than 2 years. Supplemental analyses were done to evaluate the validity of the final model shown in Table 3, particularly the following: (1) The effects of the other parameters were essentially unchanged when the ‘‘duration of ESRD prior to study start date’’ parameters were deleted from the main model, implying little to no interaction; specific interactions between ‘‘duration of ESRD prior to study start date’’ and the three nutritional parameters were tested and found to be statistically insignificant. (2) Analyses, designed to investigate the effects of mismodeling the time-dependent risks of the three nutrition-related parameters on the other factors in the Cox proportional hazards model, showed no discernible change and, in addition, supported the observation that the other factors in the model do not interact with the nutrition-related parameters. (3) The role of the caregiver classification ‘‘obese/overweight,’’ not adjusting for lnBMI, was examined in both the main Cox model (stratifying for CHF) and in two separate models based on the presence (n ⫽ 2,029, 5-year mortality 77%) or absence (n ⫽ 1578, 5-year mortality 55%) of CHF. A classifi-
1002
A
B
C
LEAVEY ET AL
Other observations from the analysis of CHF and non-CHF groups in separate models were as follows: ‘‘undernourished’’ was predictive of increased mortality risk in both groups, the effect being smaller and marginally insignificant in the non-CHF group (RR ⫽ 1.16, P ⫽ 0.08) but highly significant in the CHF group (RR ⫽ 1.35, P ⫽ 0.0047); the detrimental effect of smoking was statistically significant only in the non-CHF patients (RR ⫽ 1.25, P ⫽ 0.01); the separate models by CHF compared very closely with the reported stratified model with respect to the effects of all other explanatory variables. DISCUSSION
Most dialysis caregivers are not very familiar with the calculation or interpretation of BMI. Simply dividing standard recordings of dry weight (kg) by the square of an individual’s height (m2) (height would only need to be measured once after reaching full adult height) allows this variable to be measured. Thus, BMI adjusts weight for height to give a standardized measure of nonskeletal mass. It correlates with body ‘‘fatness’’ or ‘‘density.’’15 As might be predicted by the high prevalence of protein-energy malnutrition in ESRD patients treated with hemodialysis, a comparison with NHANES III data showed clearly that adult patients on hemodialysis are, on average, less fat than their counterparts in the general population. BMI was shown here to be of value in predicting mortality risk in hemodialysis patients independent of serum albumin and recorded clinical assessment of nutritional status. All else being
Fig 1. Time-dependent relative mortalitiy risk for (A) serum albumin, for (B) In BMI; and for (C) undernourishment.
cation of ‘‘obese’’ appeared to be overall protective, but statistically insignificant, in the stratified model (RR ⫽ 0.96, P ⫽0.44) and in the non-CHF model (RR ⫽ 0.88, P ⫽ 0.10). However, in the CHF model, ‘‘obese’’ was associated with an insignificant small increase in risk (RR ⫽ 1.04, P ⫽ 0.58).
Fig 2. Relative mortality risk for body mass index.
NUTRITION AND MORTALITY IN HEMODIALYSIS
equal, it was found that hemodialysis patients with low BMIs had a clinically significant increased risk of mortality, whereas incremental changes in BMI appeared to have a favorable effect on future well-being. In the past, the Diaphane collaborative study group has reported that overall mortality risk decreased with increasing BMI.5 Notably, however, this was reported for a cohort of younger, mostly nondiabetic, French patients treated with maintenance hemodialysis during the 1970s, and adjustment was only made for age and sex. In contrast, the data here control for a variety of other covariates in addition to age and sex and are taken from a national random sample of US patients with ESRD treated with hemodialysis in the 1990s. To generalize from the results of the Diaphane collaborative study to the current US ESRD hemodialysis population would be impossible. So, although the results of the Diaphane study are broadly consistent with this one with respect to conclusions regarding the relationship between BMI and mortality, the fact remains, that without the current study, little could be said about the relationship between BMI and mortality in the modern US ESRD hemodialysis population. The effect of a high BMI on mortality risk in hemodialysis patients was found to be in the opposite direction to that reported for healthy adults.22,23 A plausible explanation for the increased risk of death in healthy adults is that those with higher BMIs may develop hypertensive, diabetic, and hypercholesterolemic sequelae of obesity with subsequent greater mortality risk from vascular diseases. The already high prevalence of diabetes, vascular disease, and hypertension in ESRD patients at the onset of maintenance hemodialysis and high expected future incidence of vascular disease and hypertension may attenuate any negative impact of obesity on survival. Instead, it was believed plausible that greater energy reserves in hemodialysis patients might offer a protective effect for future well-being. Previously reported effects of changes in percent ideal body weight on mortality in hemodialysis patients support this hypothesis.24,25 In this mortality analysis, however, we were able to adjust not just for diabetes, demographic, and biochemical variables but also for ambulatory status, coronary heart disease, peripheral vascular disease, and left ventricular hypertrophy (a
1003
marker of significant hypertension). There were no significant interactions identified between BMI and any of these variables. This lends still further support to the hypothesis that at least in ESRD patients treated with maintenance hemodialysis, the effect of a higher BMI is to reduce mortality risk. Serum albumin was confirmed in this investigation as an independent predictor of mortality risk in hemodialysis patients. The magnitude of the RR for mortality associated with different levels in baseline serum albumin reported here is consistent with previous reports from the USRDS.10 The RR is considerably less than that initially reported by Lowrie and Lew7,8 and somewhat less than that subsequently reported by them and other groups.11,12,26 Of importance is that all studies report increasing mortality risk even when serum albumin concentrations are only slightly lower than normal. The discrepancies in the magnitude of RR described in these studies may have arisen from differences in study design such as the analysis methodology used, the use of serum albumin measured only at baseline or repeatedly over time, measurement variability in multiple versus single laboratories, and the case-mix factors used for adjustment of RR. A low serum albumin at baseline is considered to reflect diminished visceral protein stores and protein malnutrition. Malnutrition may predispose hemodialysis patients to infectious complications that cause a large proportion of deaths in this population.9,27-30 Conversely, a low serum albumin also may occur secondary to acute or chronic infectious and inflammatory processes, which in turn may have an effect on mortality risk.31-35 In any event, prognostic implications of a low serum albumin are clear. The caregiver classification ‘‘undernourished’’ may have reflected knowledge of any number of additional variables, including dietary history, gastrointestinal and other complicating medical or surgical illnesses, and psychosocial and socioeconomic factors. Although a highly subjective variable, it permitted demonstration of the value of caregivers’ observations, over and above the other measured nutritional variables. The diminishing predictive value with time of the caregiver classification ‘‘undernourished’’ may possibly reflect early mortality of the most severely malnourished patients. In all analyses examined, not
1004
adjusting for BMI, the caregiver classification ‘‘obese’’ conferred a nonsignificant protective effect, except in patients with CHF, where ‘‘obese’’ was nonsignificantly associated with an increased mortality risk. In general, we suspect that the subjective measurement ‘‘obese’’ is a less reliable measure than the objective measurement BMI. An important observation of this study was the lack of a discernible effect of ‘‘duration of ESRD prior to study start date’’ on the RR for mortality associated with the three nutritional parameters. Neither testing ‘‘duration of ESRD prior to study start date’’ as an interaction variable, nor omitting it altogether, significantly influenced the effects of serum albumin, BMI, or ‘‘undernourished’’ in the models tested. This does not necessarily imply that duration of hemodialysis does not contribute to progressively worsening nutritional status for many patients, or that serum albumin, BMI, or the classification ‘‘undernourished’’ are not affected by the duration of time on hemodialysis. Clearly sicker, more poorly nourished patients may die, and well-nourished patients may become progressively malnourished with time spent on hemodialysis. Instead, the results of this analysis suggest that markers of malnutrition have similar predictive values for future mortality risk in individuals with ESRD treated with hemodialysis, irrespective of any given individual’s duration of ESRD at the point these markers are measured; put another way, it implies that nutrition is important for all ESRD hemodialysis patients, incident or prevalent. Multiple serious complications occur for many patients during the course of ESRD treatment over a 5-year period. Against this background, the gradual decrease over time in the RR associated with baseline measurements of serum albumin, BMI, and the caregiver classification ‘‘undernourished’’ was not surprising. A Cox proportional hazards model assumes that the effect of an explanatory variable on mortality risk is constant over time. The time-dependent effects of serum albumin, BMI, and ‘‘undernourished’’ in this analysis thus violated this assumption. A shorter follow-up might have allowed a standard Cox proportional hazards model to adequately describe the RRs of the nutritional variables, but at the same time there would have been substantially fewer deaths and hence greater
LEAVEY ET AL
variability. In addition, this would have compromised the ability to describe the long-term effects of baseline nutritional parameters. As an aid to interpretation, it is suggested that, in this analysis, the RRs predicted for the three nutritional parameters by the Cox model in Table 3 be viewed as analogous to ‘‘time-weighted’’ averages. The time-dependent variation in the mortality risk associated with baseline measurement of these variables is described by timedependent coefficient models, which include the same case-mix controls as the Cox model. Timedependent coefficient models allow a single covariate measurement to act differently (ie, nonproportionally) on mortality risk according to the amount of time that has elapsed since the measurement was taken. Thus, one can evaluate the long-term effect that the baseline level of a variable has on future mortality risk. A baseline measurement (in this study, serum albumin and BMI) that has predictive value far out into the future provides useful information; in essence, it says that even after averaging all possible trajectory measurements that a given covariate might take after baseline, the initial measurement retains importance in predicting one’s future mortality risk. A baseline measurement having less ability to predict mortality far out into the future (in this study, ‘‘undernourished’’) suggests that distribution of the covariate may have changed substantially over time relative to its distribution at baseline. In situations in which a variable changes substantially over time, a time-dependent covariate model, which takes into account serial measurements, has value in describing the instantaneous effect (that is, the effect at a given time) of the variable level on a particular individual’s level of risk. This type of analysis is used in another recent paper addressing the use of timedependent variables to predict mortality for hemodialysis patients.14 It is recognized that this study cannot describe cause-and-effect relationships, but it has a number of strengths facilitating detection of significant associations between the explanatory variables examined and mortality. These include the randomized selection of a large national sample of patients; careful adjustment for case mix; follow-up data for mortality for 5 years; the large number of observed mortality events; and the multiple subanalyses used to study interactions,
NUTRITION AND MORTALITY IN HEMODIALYSIS
subgroups of patients, and time dependence. The historical prospective nature of the study design meant that variables were abstracted from historical records (⫾1 month from study start date), and their predictive value for future mortality risk was investigated. In this way, potential bias arising from selective recollection of variables was avoided. The exclusion of the small group of patients who had diagnoses of neoplasm and cirrhosis before the study start was also justifiable on methodological grounds to avoid bias. This investigation of ESRD patients treated with hemodialysis in the United States shows that serum albumin concentration, BMI, and the caregiver classification ‘‘undernourished’’ have additive value as predictors of mortality risk while adjusting for demographic covariates, comorbid conditions, and ambulatory status. A remarkable and unique observation of this study is the demonstration from time-dependent mortality models that baseline measurements of serum albumin and BMI continue to have value as mortality risk predictors as late as 5 years after initial measurement. Attention to all available nutritional parameters may serve to alert caregivers to potentially remediable physical, psychological, or social circumstances that are contributing to an overall poor nutritional state. Reducing the proportion of patients who commence dialysis with protein-energy malnutrition through improved care before reaching ESRD, or by extrapolation, preventing the development or correcting protein-energy malnutrition in established hemodialysis patients, may improve outcomes for hemodialysis patients in the future. REFERENCES 1. Slomowitz LA, Monteon FJ, Grosvenor M, Laidlaw SA, Kopple JD: Effect of energy intake on nutritional status in maintenance hemodialysis patients. Kidney Int 35:704711, 1989 2. Kopple JD: McCollum Award Lecture, 1996: Proteinenergy malnutrition in maintenance dialysis patients. Am J Clin Nutr 65:1544-1557, 1997 3. Acchiardo SR, Moore LW, Latour PA: Malnutrition as the main factor in morbidity and mortality of hemodialysis patients. Kidney Int 24:S199-S203, 1983 (suppl 16) 4. Marckmann P: Nutritional status and mortality of patients in regular dialysis therapy. J Intern Med 226:429432, 1989 5. Degoulet P, Legrain M, Reach I, Aime F, Devries C, Rojas P, Jacobs C: Mortality risk factors in patients treated by chronic hemodialysis: Report of the Diaphane collaborative study. Nephron 31:103-110, 1982
1005
6. Shapiro JI, Argy WP, Rakowski TA, Chester A, Siemsen AS, Schreiner GE: The unsuitability of BUN as a criterion for prescription dialysis. Trans Am Soc Artif Intern Organs 29:129-134, 1983 7. Lowrie EG, Lew NL: Death risk in hemodialysis patients: The predictive value of commonly measured variables and an evaluation of death rate differences between facilities. Am J Kidney Dis 15:458-482, 1990 8. Lowrie EG, Huang WH, Lew NL: Death risk predictors among peritoneal dialysis and hemodialysis patients: A preliminary comparison. Am J Kidney Dis 26:220-228, 1995 9. Churchill DN, Taylor DW, Cook RJ, LaPlante P, Barre P, Cartier P, Fay WP, Goldstein MB, Jindal K, Mandin H, McKenzie JK, Muirhead N, Parfrey PS, Posen GA, Slaughter D, Ulan RA, Werb R: Canadian Hemodialysis Morbidity Study. Am J Kidney Dis 19:214-234, 1992 10. United States Renal Data System: Co-morbid conditions and correlations with mortality risk among 3399 incident hemodialysis patients. Am J Kidney Dis 20:20-26, 1992 (suppl 2) 11. Owen WF Jr, Lew NL, Liu Y, Lowrie EG, Lazarus JM: The urea reduction ratio and serum albumin concentration as predictors of mortality in patients undergoing hemodialysis. N Engl J Med 329:1001-1006, 1993 12. Collins AJ, Ma JZ, Umen A, Keshaviah P: Urea index and other predictors of hemodialysis patient survival. Am J Kidney Dis 23:272-282, 1994 and 24:157, 1994 13. Mailloux LU, Napolitano B, Bellucci AG, Mossey RT, Vernace MA, Wilkes BM: The impact of co-morbid risk factors at the start of dialysis upon the survival of ESRD patients. ASAIO J 42:164-169, 1996 14. Culp K, Flanigan M, Lowrie EG, Lew N, Zimmerman B: Modeling mortality risk in hemodialysis patients using laboratory values as time-dependent covariates. Am J Kidney Dis 28:741-746, 1996 15. Bouchard C: Genetics of Obesity. Boca Raton, FL, CRC Press, 1994 16. Kuczmarski RJ, Flegal KM, Campbell SM, Johnson CL: Increasing prevalence of overweight among US adults: The National Health and Nutrition Examination Surveys, 1960 to 1991. JAMA 272:205-211, 1994 17. Cox DR: Regression models and life tables. J R Stat Soc, Series B 34:187-202, 1972 18. Kooperberg C, Stone C, Truang YK: Hazard Regression. J Am Stat Assoc 90:78-94, 1995 19. Gray RJ: Flexible methods for analyzing survival data using splines, with applications to breast cancer prognosis. J Am Stat Assoc 87:942-951, 1992 20. Cain KC, Lange NT: Approximate case influence for the proportional hazards regression model with censored data. Biometrics 40:493-499, 1984 21. Lin DY, Wei LJ: The robust inference for the Cox proportional hazards model. J Am Stat Assoc 84:1074-1078, 1989 22. Manson JE, Willett WC, Stampfer MJ, Colditz GA, Hunter DJ, Hankinson SE, Hennekens CH, Speizer FE: Body weight and mortality among women. N Engl J Med 333:677-685, 1995 23. Byers T: Body weight and mortality. N Engl J Med 333:723-724, 1995
1006
24. Lowrie EG, Huang WH, Lew NL, Liu Y: The relative contribution of measured variables to death risk among hemodialysis patients, in Friedman EA, (ed): Death on Hemodialysis, chap 13. the Netherlands, Kluwer, 1994, pp 121-141 25. Lowrie EG: Conceptual model for a core pathobiology of uremia with special reference to anemia, malnourishment, and mortality among dialysis patients. Semin Dial 10:115-129, 1997 26. Keane WF, Collins AJ: Influence of co-morbidity on mortality and morbidity in patients treated with hemodialysis. Am J Kidney Dis 24:1010-1018, 1994 27. Sengar DP, Rashid A, Harris JF: In vitro cellular immunity and in vivo delayed hypersensitivity in uremic patients maintained on hemodialysis. Arch Allergy Appl Immunol 47:829-838, 1974 28. Bansal VK, Popli S, Pickering J, Ing TS, Vertuno LL, Hano JE: Protein-calorie malnutrition and cutaneous anergy in hemodialysis maintained patients. Am J Clin Nutr 33:16081611, 1980 29. Mattern WD, Hak LJ, Lamanna RW, Teasley KM, Laffell MS: Malnutrition, altered immune function, and the risk of infection in maintenance hemodialysis patients. Am J Kidney Dis 1:206-218, 1982
LEAVEY ET AL
30. United States Renal Data System: USRDS 1996 Annual Data Report, Bethesda, MD, National Institutes of Health, National Institute of Diabetes and Digestive and Kidney Diseases, 1996, and Am J Kidney Dis 28:S1-S166, 1996 (suppl 2) 31. Kaysen GA, Stevenson FT, Depner TA: Determinants of albumin concentration in hemodialysis patients. Am J Kidney Dis 29:658-668, 1997 32. Barany P, Divino-Filho JC, Bergstrom J: High C-reactive protein is a strong predictor of resistance to erythropoeitin in hemodialysis patients. Am J Kidney Dis 29:565-568, 1997 33. Kaysen GA, Rathore V, Shearer GC, Depner TA: Mechanisms of hypoalbuminemia in hemodialysis patients. Kidney Int 48:510-516, 1995 34. Qureshi AR, Anderstam B, Danielson A, Gutierrez B, Lindholm B, Bergstrom J: Predictors of malnutrition in maintenance hemodialysis (HD) patients. J Am Soc Nephrol 6:586, 1995 (abstr) 35. Bergstrom J, Heimburger O, Lindholm B, Qureshi AR: Elevated serum C-reactive protein is a strong predictor of increased mortality and low serum albumin in hemodialysis (HD) patients. J Am Soc Nephrol 6:573, 1995 (abstr)
A
HIGH INCIDENCE of mortality complicates the treatment of hemodialysis patients with end-stage renal disease (ESRD). Protein and caloric intakes in maintenance hemodialysis patients have been observed to fall, on average, below the recommended requirements,1,2 and protein-energy malnutrition is considered to be a major determinant of morbidity and mortality.3-14 Several studies have documented a strong relationship between serum albumin concentration and mortality.7,8,10-14 These observations have been in both prevalent and incident ESRD patients. Serum albumin is considered to reflect visceral protein stores and thus to act as a marker of protein malnutrition. Energy malnutrition is more difficult to estimate, and whether it independently predicts an increased mortality risk in maintenance hemodialysis patients is unclear. The anthropometric measurement body mass index (BMI), calculated as an individual’s weight in kilograms divided by the square of their height in meters (kg/m2), correlates sufficiently with direct measures of body
fatness15 (eg, as measured by hydrodensitometry), to be used as a marker of energy nutritional status (energy storage). Data collected by the United States Renal Data System (USRDS) Case Mix Adequacy study, on a randomly selected national sample of Medicare ESRD patients who were treated with maintenance hemodialysis, provided a unique opportunity to examine the distribution and independent predictive value for mortality, controlling for case mix (eg, age, race, gender, and various comorbidities) of baseline measures of BMI, serum albumin concentration,
From the United States Renal Data System and the Kidney Epidemiology and Cost Center, Departments of Internal Medicine and Biostatistics, University of Michigan, Ann Arbor, MI and Bethesda, MD. Received September 30, 1997; accepted in revised form December 19, 1997. Address reprint requests to Sean F. Leavey, MB, MRCP, Kidney Epidemiology and Cost Center, 315 W Huron St, Suite 240, Ann Arbor, MI 48103. E-mail: [email protected]
r 1998 by the National Kidney Foundation, Inc. 0272-6386/98/3106-0014$3.00/0
American Journal of Kidney Diseases, Vol 31, No 6 (June), 1998: pp 997-1006
997
998
LEAVEY ET AL
and a third nutritional indicator, the nutritional status as documented in the medical records, by the clinician/caregiver. All three nutrition-related parameters, measured at baseline, were shown in this study to independently predict mortality risk. The effects of baseline BMI and serum albumin remained important up to 5 years after their date of measurement.
Table 1. Demographic Characteristics and Distributions of Baseline Comorbid Covariate Values for the Overall Random Sample Population (n ⴝ 6,585) and the Survival Analysis Subpopulation (n ⴝ 3,607)
Variable*
MATERIALS AND METHODS
Data The data used in this analysis were collected for the USRDS Case Mix Adequacy study and were supplemented with additional mortality follow-up data available from the USRDS patient database. The Case Mix Adequacy study collected data on a random sample of Medicare patients who were alive on hemodialysis on the study start date (December 31, 1990). The ‘‘duration of ESRD prior to study start date’’ varied significantly within the study population and was accounted for in all analyses. The data were abstracted retrospectively from the patients’ dialysis facility medical records by the 18 ESRD Networks under a contract with the Health Care Financing Administration (HCFA) of the U.S. Department of Health and Human Services. A copy of the data abstraction form for this study is published in the USRDS 1996 Annual Data Report, Appendix B. Details abstracted from each patient’s dialysis facility medical record included patient demographics; the presence or absence of a variety of comorbid health conditions occurring within 10 years before the study start date (Table 1); patient height (recorded any time); patient dry weight; serum albumin concentration; and documented nutritional status. The dry weight and documented caregiver classification of nutritional status were obtained from ⫾1 month of the study start date (December 1990 and January 1991). If dry weight was not clearly documented within that period, then the lowest weight within ⫾2 weeks of the study start date was recorded as the dry weight. BMI was computed from an individual’s dry weight in kilograms divided by the square of their height in meters (kg/m2). The serum albumin concentration was taken as the average of at least two values obtained within ⫾1 month of the study start date. Supplemental information on mortality through December 31, 1995, were obtained from the main USRDS patient database. To compare the relative size (as measured through BMI) of ESRD patients with their healthier counterparts in the general US population, we obtained data on BMI for the latter from the National Health and Nutrition Examination Survey III (NHANES III) phase 1 study.16 NHANES III, the third in a series of cross-sectional surveys conducted by the National Center for Health Statistics, was conducted between 1988 and 1994. It examined a nationally representative sample of the US civilian noninstitutionalized population and is an excellent source of information on a variety of health-related measures, including BMI.
Demographics Mean age, years (SD) Race (% black) Gender (% female) Vintage, years (SD)† Nutritional parameters Mean BMI, kg/m2 (SD) Mean serum albumin, g/dL (SD) ‘‘Undernourished’’ (caregiver classification‡) % ‘‘Obese’’ (caregiver classification‡) % Comorbid conditions (% yes) Diabetes Coronary artery disease§ Congestive heart failure Left ventricular hypertrophy Peripheral vascular disease\ Smoking Pulmonary disease¶ Functional status (% yes) Unable to ambulate independently
Full Database (Max n ⫽ 6,585) Mean (SD) or Percent
Survival Analysis (n ⫽ 3,607) Mean (SD) or Percent
58.5 (16) 35.9 49.6 3.4 (3.8)
57.1 (16) 37.1 50.0 3.4 (3.8)
24.3 (5.8)
24.4 (5.3)
3.7 (0.5)
3.7 (0.5)
22.2
20.1
24.3
25.0
40.0 46.0 43.2
41.2 47.7 43.7
39.0
40.9
22.2 16.8 11.4
23.0 17.1 11.6
8.3
7.6
*All of these variables are included in the survival analysis. †Vintage refers to the mean duration of ESRD before study start date. ‡Includes documentation in the medical records of ‘‘undernourished/cachectic,’’ ‘‘obese/overweight,’’ or ‘‘well nourished’’ within 1 month of study start date. §Includes history of angina, myocardial infarction, coronary angioplasty, coronary artery bypass surgery, or abnormal coronary angiography. \Includes history of claudication, amputation, absent foot pulses, and prior diagnosis of peripheral vascular disease. ¶Includes chronic obstructive pulmonary disease, asthma, and use of home oxygen.
Analyses Baseline distributions for the three major nutritional parameters and several important indicators of case-mix severity were determined for the total possible sample population in the Case Mix Adequacy study (n ⫽ 6,585). Comparison
NUTRITION AND MORTALITY IN HEMODIALYSIS
was made with their respective distributions for the 3,607 hemodialysis patients who were included in the current mortality analyses, and who had no missing values for any of the nutritional parameters. A one-sided, two-sample t-test was used to compare age-specific mean BMI for men and women in the Case Mix Adequacy population to the mean BMI for men and women in the general US population 20 years of age or older. Missing values for the three major nutritional parameters primarily accounted for the decreased sample size used in the mortality analyses as follows: 1,140 patients were missing BMI (mostly because of missing height); 1,696 patients were missing documentation of their nutritional status within 1 month of December 31, 1990; and 293 patients were missing documentation of serum albumin levels. In addition, 636 patients with neoplasm and 116 patients with cirrhosis were also excluded from the analysis, because it was thought that including both of these factors might adversely impact a fair evaluation of the nutrition-related measures on mortality. Exclusion for one or more of these reasons left a final sample size of 3,607 patients. The distribution of BMI in our sample was found to be highly skewed, with a long, right-hand tail. Hazard regression models, including but not limited to Cox proportional hazards regression, were used to evaluate the relationship between BMI and mortality in this study.17-19 It can be shown mathematically that the influence of any single observation on the estimated regression coefficient in most regression models depends on the distance of that individual’s covariate value from the sample mean value; such is the case, for example, in Cox proportional hazards regression.20 Consequently, to avoid a situation in which a few high BMI observations could easily have had an undue influence on the results, we used a natural log transformation of BMI (lnBMI); the latter was essentially normally distributed about its mean value. In the mortality analyses, all continuous variables (save ‘‘duration of ESRD prior to study start date’’) were centered at their mean values; hence, we measured unit changes in lnBMI, serum albumin, and age (1.0 units, 1 g/dL, 1 year, respectively) relative to their respective mean values (ie, relative risk [RR] ⫽ 1.0 at the mean). Exploratory mortality analyses, performed for the purposes of model building and proportional hazard assessments, were done using both HARE (a nonparametric hazard regression program) and time-dependent coefficient hazard regression models.18,19 Using HARE, no significant evidence of nonlinear effects on the RR for mortality was found for any of the continuous variables used in the mortality analysis, save ‘‘duration of ESRD prior to study start date.’’ The increase in the RR for mortality associated with each year increase in the ‘‘duration of ESRD prior to study start date’’ slowed substantially after 2 years, the baseline group having a ‘‘duration of ESRD prior to study start date’’ of 0 years. A linear spline function for ‘‘duration of ESRD prior to study start date,’’ with a single breakpoint at 2 years, was used in the final reported Cox regression model to account for this change in RR. Significant interactions between age and diabetes, and black race and diabetes, were identified and are included in the final model. In addition to the HARE models, the use of timedependent coefficient models19 strongly indicated nonpropor-
999
tional hazards (non-PH) effects over time of congestive heart failure (CHF), serum albumin, ‘‘undernourished,’’ and lnBMI. Stratifying on CHF conveniently addressed the nonproportional effects of this variable without resulting in unbalanced strata. This was not an option for the latter three variables because their effects were of direct interest in this investigation. A robust variance estimate was used to help account for the fact that there were known imperfections (eg, due to non-PH) in the final reported Cox proportional hazards model.21 Plots were generated to show (1) the changes in the estimated relative mortality risk over time for serum albumin, lnBMI, and ‘‘undernourished’’ obtained under the time-dependent coefficient model19; and (2) the relationship between BMI and mortality risk (through an estimated RR for lnBMI) suggested by a Cox proportional hazards model. A sensitivity analysis, to verify the linear effect of lnBMI on lnRR, was done where BMI was categorized into quintiles, such that there were an equal number of deaths in each quintile. A Cox model was fit, replacing lnBMI with indicators for each of the lower and higher two quintiles (the middle quintile constituting the reference group, that is, those assigned an RR of 1.0). Supplemental analyses were done to evaluate the effects of the explanatory variables separately in those patients with and without CHF. Supplemental models also looked at the effects of the caregiver classification ‘‘obese.’’ An independent effect of obese and BMI was believed to be implausible, and to investigate this, we evaluated the effects of obesity on mortality risk with and without adjusting for BMI. All mortality analyses were done using the S-Plus Statistical Software System version 3.3, (StatSci, Seattle, WA). For all analyses, statistical significance is reported at the ␣ ⫽ 0.05 level.
RESULTS
The demographic characteristics and baseline comorbid covariates for both the total random sample Case Mix Adequacy study population (n ⫽ 6,585) and the subpopulation included in the mortality analysis (n ⫽ 3,607) are shown in Table 1. For all variables, the subpopulation was similar to the total population. Comparisons of BMI among male and female hemodialysis patients and the general US population, by age group, are shown in Table 2. In all age groups and both genders mean BMI was significantly lower in hemodialysis patients than in the respective general population samples. The distribution of BMI in the sample was found to be highly skewed, with a long righthand tail. The distributions of serum albumin and lnBMI, however, were approximately normally distributed. Serum albumin and lnBMI also showed a very weak correlation of r ⫽ 0.04. These results suggest that lnBMI and serum albumin essentially varied independently of one
1000
LEAVEY ET AL
Table 2. Comparison of Age-Specific Mean Body Mass Index for Hemodialysis Patients in Case Mix Adequacy Study Versus General US Population in NHANES III Phase I Study by Gender General Population*
Age (yr)
Men 20-29 30-39 40-49 50-59 60-69 70-79 ⬎80 Women 20-29 30-39 40-49 50-59 60-69 70-79 ⬎80
Hemodialysis Population
Sample Size
Mean BMI (SE)
Sample Size
Mean BMI (SE)
P
858 759 643 493 588 495 373
24.9 (0.21) 26.1 (0.29) 27.3 (0.36) 27.6 (0.16) 26.9 (0.22) 26.5 (0.29) 24.7 (0.24)
134 327 377 490 675 523 141
22.1 (0.29) 23.5 (0.26) 25.0 (0.28) 24.8 (0.23) 23.9 (0.16) 22.9 (0.16) 21.7 (0.29)
⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001
755 771 624 464 595 446 396
24.1 (0.29) 26.4 (0.39) 26.7 (0.29) 28.5 (0.41) 27.3 (0.27) 26.7 (0.29) 24.6 (0.23)
134 201 312 448 804 613 164
22.7 (0.51) 24.1 (0.42) 25.8 (0.36) 26.1 (0.27) 25.3 (0.20) 23.4 (0.20) 21.9 (0.29)
0.0103 ⬍0.0001 0.0284 ⬍0.0001 ⬍0.0001 ⬍0.0001 ⬍0.0001
*Reference 16.
another, and consequently, any statistically significant effects of lnBMI or serum albumin on mortality were unlikely to be due to the other being significant. All of the reported data on the relative risks for mortality associated with changes in any one of the three nutritional parameters, BMI, serum albumin, and ‘‘undernourished,’’ adjusted both for the effects of the other two nutritional parameters and the following case mix variables: age; race; gender; ‘‘duration of ESRD prior to study start date’’; diabetes; coronary artery disease; left ventricular hypertrophy; peripheral vascular disease; smoking status; pulmonary disease; and ambulatory status. Unless otherwise specified, the models were stratified on CHF. In all, 64.7% of patients died over the 5-year follow-up period. The results of a Cox proportional hazards analysis of mortality risk are shown in Table 3. A lower mortality risk accompanied higher values of both lnBMI and serum albumin, relative to their mean values. An increased RR for mortality accompanied the caregiver classification ‘‘undernourished.’’ The effect of each nutritional parameter on mortality risk was seen to
be highly statistically significant (P ⬍ 0.0005 in each instance) even when adjusting for the other two. The effects of each of these parameters were also found to be time dependent. The timedependent changes in RR are plotted in Figure 1A through C. For each variable, both the fixed RR from the Cox proportional hazards model of Table 3 and the time-varying RR with 95% confidence intervals estimated from the timedependent coefficient model are shown. The effect on mortality risk associated with baseline measurements of ‘‘undernourished,’’ serum albumin, and lnBMI was greatest at the beginning, approached the estimated RR obtained under the Cox proportional hazards model during the second year, and, for serum albumin and lnBMI, remained statistically significant up to 5 years after the measurement was taken. A baseline diagnosis of ‘‘undernourished’’ increased the risk of mortality in the first 3 years; thereafter, the confidence intervals cross the RR of 1.0 (Fig 1C). Figure 1A through 1C illustrates that the RRs reported in the Cox proportional hazards model for serum albumin, BMI, and ‘‘undernourished’’ (Table 3) might be best regarded as ‘‘timeweighted’’ averages. Figure 2 contains a plot of the RR for mortality as a function of BMI. The solid line in this plot shows a decreasing relative risk of mortality with an increase in BMI over the range of values seen in this study, or 15 to 45 kg/m2. All other variables being equal, this range of BMI measurements corresponds to a range in the estimated RR for mortality of 1.2 to 0.8 (RR of 1.0 being centered at the geometric mean BMI for the study population, 23.9 kg/m2). The solid line was generated from the fitted Cox model described in Table 3, which treats BMI as a continuous variable. The model fit assumes that the RR is linear (on the natural log scale) in lnBMI. Superimposed on this plot are the results of a separate sensitivity analysis. The intent of this sensitivity analysis was to evaluate whether the solid line, which is based on the above parametric assumption, reasonably represents the relationship between mortality risk and BMI. Five categories of BMI were used in this sensitivity analysis and corresponded to the quintiles of BMI among those who died. BMI was entered as a set of four categorical variables, rather than one continuous variable, into an otherwise unchanged Cox pro-
NUTRITION AND MORTALITY IN HEMODIALYSIS
1001
Table 3. Proportional Hazards Relative Mortality Estimates for Hemodialysis Patient Risk Factors Present at Study Start Date (Stratified for Congestive Heart Failure)
Demographic, Nutritional, or Comorbid Factors and Interactions
Regression Coefficient
Robust SE
Relative Mortality Rate
P
Age (per 1 year older) Race (black v nonblack) Gender (female v male) Vintage 0-2 yr (per yr) Vintage ⬎2 yr (per yr) LnBMI (per 1 unit, eg, 3.7 v 2.7)* Serum albumin (per 1.0 g/dL higher) Undernourished (yes v no) Diabetic (yes v no) Coronary artery disease (yes v no) Left ventricular hypertrophy (yes v no) Peripheral vascular disease (yes v no) Smoking, history of (yes v no) Nonambulatory (yes v no) Pulmonary disease (yes v no) Interactions: Age ⴱ diabetes (v age ⴱ nondiabetic) Black ⴱ diabetes (v black ⴱ nondiabetic)
0.0374 ⫺0.0567 ⫺0.0375 0.1272 ⫺0.1142 ⫺0.4470 ⫺0.3950 0.2223 0.5563 0.2274 0.0896 0.2959 0.1776 0.4436 0.2373
0.0022 0.0671 0.0449 0.0282 0.0318 0.1207 0.0530 0.0610 0.0636 0.0473 0.0449 0.0542 0.0637 0.0815 0.0706
1.04 0.95 0.96 1.14 0.89† 0.64 0.67 1.25 1.74 1.26 1.09 1.34 1.19 1.56 1.27
⬍0.0001 0.40 0.40 ⬍0.0001 ⬍0.0001 0.0002 ⬍0.0001 0.0003 ⬍0.0001 ⬍0.0001 0.046 ⬍0.0001 0.0053 ⬍0.0001 0.0008
⫺0.0097 ⫺0.2247
0.0033 0.0913
0.99 0.80
0.0029 0.014
*BMI 40.4 v 14.9, see Figure 2 for BMI. †RR ⫽ 1.01 per year after 2 years.
portional hazards model (see Table 3). The reference group was taken to be the middle quintile, which by definition had an RR of 1.0. A significantly higher risk for mortality was seen in the lowest quintile group (P ⫽ 0.02), and a marginally insignificant lower risk for mortality (P ⫽ 0.098) was seen in the highest quintile group. The fact that the solid line lies well within the 95% confidence limits obtained from the model treating BMI as a set of four categorical variables indicates that the solid line does indeed provide a concise and reasonable description of the relationship between mortality risk and BMI. Two significant interactions were detected: the increasing RR of mortality with increasing age, although still present, was less pronounced in diabetics than in nondiabetics; and the protective effect of being black was greater in diabetics than in nondiabetics. The main model adjusted for both of these interactions (Table 3). A spline function describes the effect of ‘‘duration of ESRD prior to study start date.’’ The increasing RR for mortality is steeper for ‘‘duration of ESRD prior to study start date’’ up to 2 years and less steep for ‘‘duration of ESRD prior to study start date’’ greater than 2 years. The main model therefore allows for differential effects on mortal-
ity for ‘‘duration of ESRD prior to study start date’’ up to 2 years and ‘‘duration of ESRD prior to study start date’’ greater than 2 years. Supplemental analyses were done to evaluate the validity of the final model shown in Table 3, particularly the following: (1) The effects of the other parameters were essentially unchanged when the ‘‘duration of ESRD prior to study start date’’ parameters were deleted from the main model, implying little to no interaction; specific interactions between ‘‘duration of ESRD prior to study start date’’ and the three nutritional parameters were tested and found to be statistically insignificant. (2) Analyses, designed to investigate the effects of mismodeling the time-dependent risks of the three nutrition-related parameters on the other factors in the Cox proportional hazards model, showed no discernible change and, in addition, supported the observation that the other factors in the model do not interact with the nutrition-related parameters. (3) The role of the caregiver classification ‘‘obese/overweight,’’ not adjusting for lnBMI, was examined in both the main Cox model (stratifying for CHF) and in two separate models based on the presence (n ⫽ 2,029, 5-year mortality 77%) or absence (n ⫽ 1578, 5-year mortality 55%) of CHF. A classifi-
1002
A
B
C
LEAVEY ET AL
Other observations from the analysis of CHF and non-CHF groups in separate models were as follows: ‘‘undernourished’’ was predictive of increased mortality risk in both groups, the effect being smaller and marginally insignificant in the non-CHF group (RR ⫽ 1.16, P ⫽ 0.08) but highly significant in the CHF group (RR ⫽ 1.35, P ⫽ 0.0047); the detrimental effect of smoking was statistically significant only in the non-CHF patients (RR ⫽ 1.25, P ⫽ 0.01); the separate models by CHF compared very closely with the reported stratified model with respect to the effects of all other explanatory variables. DISCUSSION
Most dialysis caregivers are not very familiar with the calculation or interpretation of BMI. Simply dividing standard recordings of dry weight (kg) by the square of an individual’s height (m2) (height would only need to be measured once after reaching full adult height) allows this variable to be measured. Thus, BMI adjusts weight for height to give a standardized measure of nonskeletal mass. It correlates with body ‘‘fatness’’ or ‘‘density.’’15 As might be predicted by the high prevalence of protein-energy malnutrition in ESRD patients treated with hemodialysis, a comparison with NHANES III data showed clearly that adult patients on hemodialysis are, on average, less fat than their counterparts in the general population. BMI was shown here to be of value in predicting mortality risk in hemodialysis patients independent of serum albumin and recorded clinical assessment of nutritional status. All else being
Fig 1. Time-dependent relative mortalitiy risk for (A) serum albumin, for (B) In BMI; and for (C) undernourishment.
cation of ‘‘obese’’ appeared to be overall protective, but statistically insignificant, in the stratified model (RR ⫽ 0.96, P ⫽0.44) and in the non-CHF model (RR ⫽ 0.88, P ⫽ 0.10). However, in the CHF model, ‘‘obese’’ was associated with an insignificant small increase in risk (RR ⫽ 1.04, P ⫽ 0.58).
Fig 2. Relative mortality risk for body mass index.
NUTRITION AND MORTALITY IN HEMODIALYSIS
equal, it was found that hemodialysis patients with low BMIs had a clinically significant increased risk of mortality, whereas incremental changes in BMI appeared to have a favorable effect on future well-being. In the past, the Diaphane collaborative study group has reported that overall mortality risk decreased with increasing BMI.5 Notably, however, this was reported for a cohort of younger, mostly nondiabetic, French patients treated with maintenance hemodialysis during the 1970s, and adjustment was only made for age and sex. In contrast, the data here control for a variety of other covariates in addition to age and sex and are taken from a national random sample of US patients with ESRD treated with hemodialysis in the 1990s. To generalize from the results of the Diaphane collaborative study to the current US ESRD hemodialysis population would be impossible. So, although the results of the Diaphane study are broadly consistent with this one with respect to conclusions regarding the relationship between BMI and mortality, the fact remains, that without the current study, little could be said about the relationship between BMI and mortality in the modern US ESRD hemodialysis population. The effect of a high BMI on mortality risk in hemodialysis patients was found to be in the opposite direction to that reported for healthy adults.22,23 A plausible explanation for the increased risk of death in healthy adults is that those with higher BMIs may develop hypertensive, diabetic, and hypercholesterolemic sequelae of obesity with subsequent greater mortality risk from vascular diseases. The already high prevalence of diabetes, vascular disease, and hypertension in ESRD patients at the onset of maintenance hemodialysis and high expected future incidence of vascular disease and hypertension may attenuate any negative impact of obesity on survival. Instead, it was believed plausible that greater energy reserves in hemodialysis patients might offer a protective effect for future well-being. Previously reported effects of changes in percent ideal body weight on mortality in hemodialysis patients support this hypothesis.24,25 In this mortality analysis, however, we were able to adjust not just for diabetes, demographic, and biochemical variables but also for ambulatory status, coronary heart disease, peripheral vascular disease, and left ventricular hypertrophy (a
1003
marker of significant hypertension). There were no significant interactions identified between BMI and any of these variables. This lends still further support to the hypothesis that at least in ESRD patients treated with maintenance hemodialysis, the effect of a higher BMI is to reduce mortality risk. Serum albumin was confirmed in this investigation as an independent predictor of mortality risk in hemodialysis patients. The magnitude of the RR for mortality associated with different levels in baseline serum albumin reported here is consistent with previous reports from the USRDS.10 The RR is considerably less than that initially reported by Lowrie and Lew7,8 and somewhat less than that subsequently reported by them and other groups.11,12,26 Of importance is that all studies report increasing mortality risk even when serum albumin concentrations are only slightly lower than normal. The discrepancies in the magnitude of RR described in these studies may have arisen from differences in study design such as the analysis methodology used, the use of serum albumin measured only at baseline or repeatedly over time, measurement variability in multiple versus single laboratories, and the case-mix factors used for adjustment of RR. A low serum albumin at baseline is considered to reflect diminished visceral protein stores and protein malnutrition. Malnutrition may predispose hemodialysis patients to infectious complications that cause a large proportion of deaths in this population.9,27-30 Conversely, a low serum albumin also may occur secondary to acute or chronic infectious and inflammatory processes, which in turn may have an effect on mortality risk.31-35 In any event, prognostic implications of a low serum albumin are clear. The caregiver classification ‘‘undernourished’’ may have reflected knowledge of any number of additional variables, including dietary history, gastrointestinal and other complicating medical or surgical illnesses, and psychosocial and socioeconomic factors. Although a highly subjective variable, it permitted demonstration of the value of caregivers’ observations, over and above the other measured nutritional variables. The diminishing predictive value with time of the caregiver classification ‘‘undernourished’’ may possibly reflect early mortality of the most severely malnourished patients. In all analyses examined, not
1004
adjusting for BMI, the caregiver classification ‘‘obese’’ conferred a nonsignificant protective effect, except in patients with CHF, where ‘‘obese’’ was nonsignificantly associated with an increased mortality risk. In general, we suspect that the subjective measurement ‘‘obese’’ is a less reliable measure than the objective measurement BMI. An important observation of this study was the lack of a discernible effect of ‘‘duration of ESRD prior to study start date’’ on the RR for mortality associated with the three nutritional parameters. Neither testing ‘‘duration of ESRD prior to study start date’’ as an interaction variable, nor omitting it altogether, significantly influenced the effects of serum albumin, BMI, or ‘‘undernourished’’ in the models tested. This does not necessarily imply that duration of hemodialysis does not contribute to progressively worsening nutritional status for many patients, or that serum albumin, BMI, or the classification ‘‘undernourished’’ are not affected by the duration of time on hemodialysis. Clearly sicker, more poorly nourished patients may die, and well-nourished patients may become progressively malnourished with time spent on hemodialysis. Instead, the results of this analysis suggest that markers of malnutrition have similar predictive values for future mortality risk in individuals with ESRD treated with hemodialysis, irrespective of any given individual’s duration of ESRD at the point these markers are measured; put another way, it implies that nutrition is important for all ESRD hemodialysis patients, incident or prevalent. Multiple serious complications occur for many patients during the course of ESRD treatment over a 5-year period. Against this background, the gradual decrease over time in the RR associated with baseline measurements of serum albumin, BMI, and the caregiver classification ‘‘undernourished’’ was not surprising. A Cox proportional hazards model assumes that the effect of an explanatory variable on mortality risk is constant over time. The time-dependent effects of serum albumin, BMI, and ‘‘undernourished’’ in this analysis thus violated this assumption. A shorter follow-up might have allowed a standard Cox proportional hazards model to adequately describe the RRs of the nutritional variables, but at the same time there would have been substantially fewer deaths and hence greater
LEAVEY ET AL
variability. In addition, this would have compromised the ability to describe the long-term effects of baseline nutritional parameters. As an aid to interpretation, it is suggested that, in this analysis, the RRs predicted for the three nutritional parameters by the Cox model in Table 3 be viewed as analogous to ‘‘time-weighted’’ averages. The time-dependent variation in the mortality risk associated with baseline measurement of these variables is described by timedependent coefficient models, which include the same case-mix controls as the Cox model. Timedependent coefficient models allow a single covariate measurement to act differently (ie, nonproportionally) on mortality risk according to the amount of time that has elapsed since the measurement was taken. Thus, one can evaluate the long-term effect that the baseline level of a variable has on future mortality risk. A baseline measurement (in this study, serum albumin and BMI) that has predictive value far out into the future provides useful information; in essence, it says that even after averaging all possible trajectory measurements that a given covariate might take after baseline, the initial measurement retains importance in predicting one’s future mortality risk. A baseline measurement having less ability to predict mortality far out into the future (in this study, ‘‘undernourished’’) suggests that distribution of the covariate may have changed substantially over time relative to its distribution at baseline. In situations in which a variable changes substantially over time, a time-dependent covariate model, which takes into account serial measurements, has value in describing the instantaneous effect (that is, the effect at a given time) of the variable level on a particular individual’s level of risk. This type of analysis is used in another recent paper addressing the use of timedependent variables to predict mortality for hemodialysis patients.14 It is recognized that this study cannot describe cause-and-effect relationships, but it has a number of strengths facilitating detection of significant associations between the explanatory variables examined and mortality. These include the randomized selection of a large national sample of patients; careful adjustment for case mix; follow-up data for mortality for 5 years; the large number of observed mortality events; and the multiple subanalyses used to study interactions,
NUTRITION AND MORTALITY IN HEMODIALYSIS
subgroups of patients, and time dependence. The historical prospective nature of the study design meant that variables were abstracted from historical records (⫾1 month from study start date), and their predictive value for future mortality risk was investigated. In this way, potential bias arising from selective recollection of variables was avoided. The exclusion of the small group of patients who had diagnoses of neoplasm and cirrhosis before the study start was also justifiable on methodological grounds to avoid bias. This investigation of ESRD patients treated with hemodialysis in the United States shows that serum albumin concentration, BMI, and the caregiver classification ‘‘undernourished’’ have additive value as predictors of mortality risk while adjusting for demographic covariates, comorbid conditions, and ambulatory status. A remarkable and unique observation of this study is the demonstration from time-dependent mortality models that baseline measurements of serum albumin and BMI continue to have value as mortality risk predictors as late as 5 years after initial measurement. Attention to all available nutritional parameters may serve to alert caregivers to potentially remediable physical, psychological, or social circumstances that are contributing to an overall poor nutritional state. Reducing the proportion of patients who commence dialysis with protein-energy malnutrition through improved care before reaching ESRD, or by extrapolation, preventing the development or correcting protein-energy malnutrition in established hemodialysis patients, may improve outcomes for hemodialysis patients in the future. REFERENCES 1. Slomowitz LA, Monteon FJ, Grosvenor M, Laidlaw SA, Kopple JD: Effect of energy intake on nutritional status in maintenance hemodialysis patients. Kidney Int 35:704711, 1989 2. Kopple JD: McCollum Award Lecture, 1996: Proteinenergy malnutrition in maintenance dialysis patients. Am J Clin Nutr 65:1544-1557, 1997 3. Acchiardo SR, Moore LW, Latour PA: Malnutrition as the main factor in morbidity and mortality of hemodialysis patients. Kidney Int 24:S199-S203, 1983 (suppl 16) 4. Marckmann P: Nutritional status and mortality of patients in regular dialysis therapy. J Intern Med 226:429432, 1989 5. Degoulet P, Legrain M, Reach I, Aime F, Devries C, Rojas P, Jacobs C: Mortality risk factors in patients treated by chronic hemodialysis: Report of the Diaphane collaborative study. Nephron 31:103-110, 1982
1005
6. Shapiro JI, Argy WP, Rakowski TA, Chester A, Siemsen AS, Schreiner GE: The unsuitability of BUN as a criterion for prescription dialysis. Trans Am Soc Artif Intern Organs 29:129-134, 1983 7. Lowrie EG, Lew NL: Death risk in hemodialysis patients: The predictive value of commonly measured variables and an evaluation of death rate differences between facilities. Am J Kidney Dis 15:458-482, 1990 8. Lowrie EG, Huang WH, Lew NL: Death risk predictors among peritoneal dialysis and hemodialysis patients: A preliminary comparison. Am J Kidney Dis 26:220-228, 1995 9. Churchill DN, Taylor DW, Cook RJ, LaPlante P, Barre P, Cartier P, Fay WP, Goldstein MB, Jindal K, Mandin H, McKenzie JK, Muirhead N, Parfrey PS, Posen GA, Slaughter D, Ulan RA, Werb R: Canadian Hemodialysis Morbidity Study. Am J Kidney Dis 19:214-234, 1992 10. United States Renal Data System: Co-morbid conditions and correlations with mortality risk among 3399 incident hemodialysis patients. Am J Kidney Dis 20:20-26, 1992 (suppl 2) 11. Owen WF Jr, Lew NL, Liu Y, Lowrie EG, Lazarus JM: The urea reduction ratio and serum albumin concentration as predictors of mortality in patients undergoing hemodialysis. N Engl J Med 329:1001-1006, 1993 12. Collins AJ, Ma JZ, Umen A, Keshaviah P: Urea index and other predictors of hemodialysis patient survival. Am J Kidney Dis 23:272-282, 1994 and 24:157, 1994 13. Mailloux LU, Napolitano B, Bellucci AG, Mossey RT, Vernace MA, Wilkes BM: The impact of co-morbid risk factors at the start of dialysis upon the survival of ESRD patients. ASAIO J 42:164-169, 1996 14. Culp K, Flanigan M, Lowrie EG, Lew N, Zimmerman B: Modeling mortality risk in hemodialysis patients using laboratory values as time-dependent covariates. Am J Kidney Dis 28:741-746, 1996 15. Bouchard C: Genetics of Obesity. Boca Raton, FL, CRC Press, 1994 16. Kuczmarski RJ, Flegal KM, Campbell SM, Johnson CL: Increasing prevalence of overweight among US adults: The National Health and Nutrition Examination Surveys, 1960 to 1991. JAMA 272:205-211, 1994 17. Cox DR: Regression models and life tables. J R Stat Soc, Series B 34:187-202, 1972 18. Kooperberg C, Stone C, Truang YK: Hazard Regression. J Am Stat Assoc 90:78-94, 1995 19. Gray RJ: Flexible methods for analyzing survival data using splines, with applications to breast cancer prognosis. J Am Stat Assoc 87:942-951, 1992 20. Cain KC, Lange NT: Approximate case influence for the proportional hazards regression model with censored data. Biometrics 40:493-499, 1984 21. Lin DY, Wei LJ: The robust inference for the Cox proportional hazards model. J Am Stat Assoc 84:1074-1078, 1989 22. Manson JE, Willett WC, Stampfer MJ, Colditz GA, Hunter DJ, Hankinson SE, Hennekens CH, Speizer FE: Body weight and mortality among women. N Engl J Med 333:677-685, 1995 23. Byers T: Body weight and mortality. N Engl J Med 333:723-724, 1995
1006
24. Lowrie EG, Huang WH, Lew NL, Liu Y: The relative contribution of measured variables to death risk among hemodialysis patients, in Friedman EA, (ed): Death on Hemodialysis, chap 13. the Netherlands, Kluwer, 1994, pp 121-141 25. Lowrie EG: Conceptual model for a core pathobiology of uremia with special reference to anemia, malnourishment, and mortality among dialysis patients. Semin Dial 10:115-129, 1997 26. Keane WF, Collins AJ: Influence of co-morbidity on mortality and morbidity in patients treated with hemodialysis. Am J Kidney Dis 24:1010-1018, 1994 27. Sengar DP, Rashid A, Harris JF: In vitro cellular immunity and in vivo delayed hypersensitivity in uremic patients maintained on hemodialysis. Arch Allergy Appl Immunol 47:829-838, 1974 28. Bansal VK, Popli S, Pickering J, Ing TS, Vertuno LL, Hano JE: Protein-calorie malnutrition and cutaneous anergy in hemodialysis maintained patients. Am J Clin Nutr 33:16081611, 1980 29. Mattern WD, Hak LJ, Lamanna RW, Teasley KM, Laffell MS: Malnutrition, altered immune function, and the risk of infection in maintenance hemodialysis patients. Am J Kidney Dis 1:206-218, 1982
LEAVEY ET AL
30. United States Renal Data System: USRDS 1996 Annual Data Report, Bethesda, MD, National Institutes of Health, National Institute of Diabetes and Digestive and Kidney Diseases, 1996, and Am J Kidney Dis 28:S1-S166, 1996 (suppl 2) 31. Kaysen GA, Stevenson FT, Depner TA: Determinants of albumin concentration in hemodialysis patients. Am J Kidney Dis 29:658-668, 1997 32. Barany P, Divino-Filho JC, Bergstrom J: High C-reactive protein is a strong predictor of resistance to erythropoeitin in hemodialysis patients. Am J Kidney Dis 29:565-568, 1997 33. Kaysen GA, Rathore V, Shearer GC, Depner TA: Mechanisms of hypoalbuminemia in hemodialysis patients. Kidney Int 48:510-516, 1995 34. Qureshi AR, Anderstam B, Danielson A, Gutierrez B, Lindholm B, Bergstrom J: Predictors of malnutrition in maintenance hemodialysis (HD) patients. J Am Soc Nephrol 6:586, 1995 (abstr) 35. Bergstrom J, Heimburger O, Lindholm B, Qureshi AR: Elevated serum C-reactive protein is a strong predictor of increased mortality and low serum albumin in hemodialysis (HD) patients. J Am Soc Nephrol 6:573, 1995 (abstr)