Importance of Simultaneous Evaluation of Multiple Risk Factors for Hemodialysis Patients’ Mortality and Development of a Novel Index: Dialysis Outcomes and Practice Patterns Study

Background For hemodialysis (HD) patients, many risk factors for death are associated with each other intricately. However, they are often considered separately in clinical settings. We evaluated the maintenance HD patients’ risk of death within one year from multiple risk factors simultaneously considering their interrelationships using a novel index (survival index, SI) for HD patients in the United States developed using data from the Dialysis Outcomes and Practice Patterns Study (DOPPS). Methods We analyzed data from 3899 and 3765 patients to develop and validate SI, respectively. To predict death within one year, candidate models were developed using logistic regression models. The final model was determined by comparing the accuracy among the models for the prediction of deaths. Results The model included age; body mass index; serum creatinine, albumin, total cholesterol and phosphorus levels; history of cardiovascular diseases; and arteriovenous fistula use. SI showed a higher accuracy in predicting death (c-statistic, 0.739) than geriatric nutritional risk index (0.647) and serum albumin level (0.637). The probability of death predicted on the basis of SI matched the observed number of deaths. Cox proportional hazard models for time-dependent SI showed that patients with low SI had a higher risk of death than patients with high SI [reference, Group 4 (26.1≤SI)]; Group 1 (SI<12.7), adjusted hazard ratio, 7.97 (95% CI, 5.02, 12.65); Group 2 (12.7≤SI<19.0), 3.18 (95% CI, 1.96, 5.16); Group 3 (19.0≤SI<26.1), 2.20 (95% CI, 1.33, 3.66). Conclusion Results of this study suggest that the simultaneous evaluation of multiple risk factors can more accurately assess patients’ prognosis and identify patients at an increased risk of death than single factors.


Introduction
The risk of death is high in end-stage renal disease (ESRD) patients [1]. Various risk factors for death in hemodialysis (HD) patients have been reported such as age, serum albumin and phosphorus levels, and vascular access [2]. Risk factors are intricately associated with each other. For example, the pathophysiology of malnutrition is characterized by protein-energy wasting (PEW), which leads to atherosclerotic cardiovascular diseases (CVDs), and is a strong risk factor for increased hospitalization and mortality rates in ESRD patients [3][4][5][6].
HD patients have some specific characteristics. Serum albumin level is affected by inflammation and age, and does not accurately reflect nutritional status [7]. The association of increased body mass index (BMI) with improved prognosis in HD patients is an example of "reverse epidemiology" [3,8,9], and this relationship is modified by HD patients' characteristics such as age, inflammation, and comorbid conditions [10,11]. These characteristics make it difficult to evaluate HD patients' prognosis.
Risk factors in HD patients have often been evaluated separately in clinical settings. To control all risk factors and their complex interrelationships, it is necessary to evaluate them simultaneously. However, to the best of our knowledge, there have been few studies in which the numerically integrated multiple risk factors were evaluated. The purpose of this study was to evaluate the maintenance HD patients' risk of death within one year considering multiple risk factors and their complex interrelationships. For this purpose, (1) after detection of main risk factors for death within one year using data of HD patients in the Dialysis Outcomes and Practice Patterns Study (DOPPS), a novel index (survival index, SI) was developed. (2) The accuracy of the index to predict death was compared with those of other indices. (3) High-risk patients were identified and their prognosis was evaluated.

Data Source
DOPPS was a worldwide cohort study of in-center HD practice patterns and outcomes in seven countries. The details of the method of DOPPS were reported previously [12,13]. In brief, there have been phases of DOPPS data collection since the study initiation in 1996. A randomized, stratified selection method was used to identify facilities for participation within each country. Demographic data, cause of ESRD, and mortality data for all HD patients in each facility were collected as cumulative HD census. Detailed patient data within each facility were collected at study entry and at 4-month intervals throughout the study. Institutional review boards approved DOPPS in each facility. Participants gave their informed consent to the use of their clinical records in DOPPS in accordance with the requirements of each review board and where BMI is in kg/m 2 ; Alb = serum albumin level (g/dl); height is in m 2 . A modified combined score (CS) was calculated using the following formula [19]: where a low serum albumin level of <3.5g/dl scored Yes = 1, No = 0; a low BMI of <19.6kg/m 2 was scored Yes = 1, No = 0. Although serum C-related protein (CRP) level was used in the original CS, because there were many missing values of serum CRP levels in our dataset, elevated WBC count was alternatively used in this study. The outcome was death including all-cause death and cause-specific death (CVD-caused and infection-caused death) within one year. The number of CVD-caused deaths was obtained by summing the number of patients who died of cardiac and cerebral causes. Causes of cardiac death included myocardial infarction, atherosclerotic heart disease, congestive heart failure, cardiomyopathy, arrhythmia, valvular heart disease, pericarditis, and unknown-cause cardiac arrest. Suspected PEW was diagnosed on the basis of International Society of Renal Nutrition & Metabolism (ISRNM) PEW criteria; (1) serum chemistry (serum albumin level <3.8g/dl or serum cholesterol level <100mg/dl), (2) body mass (BMI <23), and (3) dietary intake (nPCR <0.8g/kg/day) [3].

Statistical methods
Normally distributed variables are presented as mean ± standard deviation (SD); otherwise, the median and interquartile range are presented. Highly skewed variables (e.g., PTH, vintage, and CS) were transformed with the natural logarithm function prior to use in models [ln(PTH), ln (vintage), ln(CS)]. Intergroup comparisons were performed using the chi-square test, t-test, Mann-Whitney U-test, one-way analysis of variance, and Kruskal-Wallis test as appropriate.
In the dataset for the development of SI, candidate models for SI were developed using multivariate logistic regression models as follows:where P(X) = probability of all-cause death, The coefficients were rounded. Candidate variables were selected for inclusion in the SI model based on previous reports, and the results of analysis using univariate logistic regression models (p<0.1) [10,11,20]. The candidate variables were as follows: age; gender; BMI; square of BMI; race; CVDs; hypertension; DM; serum creatinine, albumin, total cholesterol and phosphorus levels; square of serum phosphorus level; elevated WBC count; controlled hemoglobin level; nPCR; AVF use; ln(vintage); and Kt/V. The square of BMI and serum phosphorus level were listed as candidate variables for evaluating the U-shaped relationship between each variable and HD patients' mortality. The interactions among BMI, the square of BMI, age, and elevated WBC count were evaluated [10,11]. Albumin-corrected calcium level and ln(PTH) were not included in the models, because there were their collinearities with serum phosphorus level, and serum phosphorus level is more associated with all-cause mortality than serum calcium and PTH levels in a systematic review [21]. Candidate models were constructed using the hierarchical backward elimination procedure: (1) the initial model was constructed as hierarchically well formulated, and (2) the interaction variables and confounders were statistically evaluated (p<0.05). Next, we constructed logistic regression models derived from the initial model by decreasing the variables to find a model that was more easily calculable than the initial model. Considering their c-statistics, Akaike information criterion (AIC), and the results of Hosmer-Lemeshow (HL) goodness-of-fit tests, several models were selected as candidate models for SI. Then, in a dataset for the SI validation, the final model was determined using the models' c-statistics, AICs, and the result of HL tests for the prediction of all-cause, CVDcaused, and infection-caused deaths within one year. The receiver operating characteristic (ROC) curves of SI for the prediction of all-cause death within one year were used to compare with those of other indices. To investigate the relationship between SI and various patient characteristic factors, Pearson's correlation coefficients were examined. The predicted probability of all-cause death occurring in each patient was predicted using the following formula: The patients were divided into four groups on the basis of the quartile cut points of SI. Patients' survival curves were derived by Kaplan-Meier analysis. Cox proportional hazards models (PHMs) were used to compare the risk of the outcome for baseline SI between groups. Then, SI was treated as a time-dependent variable. Cox PHMs for treating SI as a time-dependent covariate were also examined. Adjusted variables in Cox PHMs and time-dependent Cox PHMs were the baseline characteristics including gender, race, hypertension, DM, serum albumin-corrected calcium levels, hemoglobin level, ln(PTH), elevated WBC count, nPCR, ln(vintage), and Kt/V. The results are presented here as hazard ratios (HR) with 95% confidence interval (CI). To keep the loss of data minimum, if a patient's SI just before the primary outcome was missing, the last SI was used in the time-dependent Cox PHMs. Statistical significance was defined as p<0.05. These analyses were conducted using SAS, version 9.2 (SAS, Inc., Cary, North Carolina).

Baseline characteristics
The study population consisted of 7664 patients (Fig 1). They were randomly assigned into two groups; one group for SI development and the other for SI validation (Fig 1). No significant differences in patient characteristics between the development and validation datasets were observed ( Table 1).

Development of candidate models for SI
The initial model did not contain the following variables: gender (p = 0.54), race (p = 0.17), DM (p = 0.94), controlled hemoglobin level (p = 0.69), nPCR (p = 0.30), ln(vintage) (p = 0.27) and Kt/V (p = 0.22). Elevated WBC count; the squares of BMI and serum phosphorus level; and the variables that represented interactions were evaluated, but not retained in Model 1. We evaluated additional models containing subsets of the covariates retained in Model 1. Models 1, 3, 6, and 11 were selected as candidate models for SI, because their c-statistics for the prediction of all-cause death were higher than other models ( Table 2).
Using the dataset for the SI validation, we compared the capability of candidate models to predict all-cause death within one year. Model 3 showed the highest c-statistics for all-cause death and CVD-caused death (Table 3). Model 3 also showed an adequate fit to the data, as determined by the HL test for all-cause death. Therefore, Model 3 was selected for SI. The formula for SI was as follows: where Age is in years; BMI is in kg/m 2 ; Cr = serum creatinine level (mg/dl); Alb = serum albumin level (g/dl); Tchol = serum total cholesterol level (mg/dl); P = serum phosphorus level  2). Age showed higher c-statistics for all-cause death and CVD-caused death than GNRI, CS, BMI, serum creatinine and albumin levels. Serum creatinine and albumin levels showed higher c-statistics for hospitalization than GNRI, CS, age and BMI. BMI showed higher c-statistics for PEW than GNRI, CS, age, serum creatinine and albumin levels.
The c-statistics of SI for all-cause death were compared with those of other indices on the basis of the stratification of patients' characteristics. The c-statistics of SI in incident and maintenance HD patients were 0.743 and 0.730, respectively, and higher than other indices: GNRI

Risk of death and SI
In the validation dataset, significant differences in the distribution of baseline characteristics were observed between the groups: Group 1, SI<12.7; Group 2, 12.7 SI<19.0; Group 3, 19.0 SI<26.1; and Group 4, 26.1 SI (Table 6). In Group 1, the number of patients who died was higher than those in the other groups.
Kaplan-Meier analysis showed that the Group 1 had a higher mortality rate than the other groups [log-rank and Wilcoxon tests, p = 0.0001 (Fig 5)]. Cox PHMs and adjusted Cox PHMs showed that a 1 unit increase in SI decreased the risk of death, and that Group 1 showed a high risk of death (reference, Group 4). Analysis of SI as a time-updated effect also showed similar tends (Table 7).

Discussion
We were able to develop SI after detection of main risk factors for death within one year, and evaluate precisely the HD patients' prognosis. SI included various risk factors for death including their interrelationships, which were consistent with previous reports [2,3,22]. SI showed a higher accuracy in identifying high-risk patients than other single indices. High-risk patients were identified from the view point of multiple risk factors using SI.
The prediction of all-cause death of HD patients based on SI was more accurate than that based on SI's individual components. It has been reported that the combination of low BMI and low serum albumin and creatinine levels reflects a high risk of death more accurately than each factor [22]. On the other hand, the combination of serum albumin and CRP levels and BMI predicts HD patients' mortality more accurately than these indices [19]. These findings suggest that an index including multiple variables more accurately predicts HD patients' mortality than single indices. It has been reported that achieving multiple therapeutic targets is associated with better survival in HD patients than achieving fewer targets [23,24]. These studies indicate the importance of an index for simultaneous evaluation of multiple risk factors for HD patients' mortality.
SI was developed as a modified logit on the basis of logistic regression models. This gave unique characteristics to SI similarly to outcome-propensity score [25]. Observational studies have many confounding background characteristics. To control various confounders simultaneously and evaluate them easily, it is necessary to replace the collection of confounders with one variable. SI summarizes various risk factors for death as confounders to a single adjusted composite risk factor. The association between exposures and outcomes can be adjusted for SI in place of individual risk factors. In this study, SI showed high accuracies of prediction of all-cause death of HD patients and PEW diagnosis. SI might be appropriate as a summary index for studies in which the effects of multiple exposures on these outcomes (i.e., all-cause death and PEW) are evaluated.
Determination of HD patients' prognosis is effective for screening patients with a high risk of death and for evaluating patients' condition. The capability of an index to predict patients' risk of death clarifies which patients urgently need intervention. In this study, the predicted probability of death based on SI showed good agreement with the rate of observed all-cause death, and SI was associated with mortality in time-dependent Cox PHMs. Because SI is based on the prognosis of HD patients, can be easily measured, does not depend on the skill of examiners, and reflects HD patients' specific characteristics, it may be a useful tool to identify highrisk HD patients and evaluate their condition. In clinical settings, after the screening, the patients need to be examined in detail.
CVDs are the leading causes of death in HD patients. From the Choices for Healthy Outcomes in Caring for ESRD study, a large percentage of incident dialysis patients were found to have the common risk factors for CVDs [26]. In HD patients, a lower BMI was not a predictor of incident CVDs but an independent risk factor for death after CVDs [4,5]. PEW is one of the most important risk factors for death after CVDs [5]. SI included several factors (serum albumin, creatinine, and total cholesterol levels and BMI) listed in the ISRNM PEW criteria [3] and showed a high accuracy of PEW diagnosis. The coefficients of the variables in SI suggest that their changes toward PEW increased the risk of death. SI included serum phosphorus level. An elevated serum phosphorus level is associated with CVDs and increased mortality [20]. An increase in dietary protein intake has been shown to correlate with an increase in serum phosphorus level [27]. However, protein-intake restriction may lead to PEW. It has been reported that a prescribed reduction in phosphorus intake correlates with an increase in mortality [28]. And a decreased serum phosphorus level is also associated with increased mortality [20]. These results indicate the difficulty in keeping the balance between decreased serum phosphorus level and adequate nutrition by controlling excessive dietary intake.
There are specific associations between BMI and mortality in HD patients. First, BMI does not accurately reflect the body composition in HD patients. The protective effect of high BMI against mortality is limited to those patients with normal or high muscle mass [29]. The estimated rate of creatinine production has been used to assess the lean body mass in stable HD patients [30,31]. Both BMI and serum creatinine level in SI can be used for the patients with an imbalance between BMI and muscle mass. Moreover, reverse epidemiology in HD patients has been reported to be a phenomenon due to the effects of age and inflammation on BMI [10,11]. To adjust the interrelationships among BMI, age, and inflammation, the interactions among BMI, the square of BMI, age, and inflammation were evaluated, and finally BMI and age were included in SI. The coefficients of BMI and age in SI suggest that the patients' prognosis worsened with decreasing in BMI and aging.
It has been reported that the relationship between BMI and mortality is U-shaped among HD patients aged <65 years and almost linear among those aged !65 years [11]. With the development of SI, it was observed that the coefficients of variables such as BMI and the square of BMI in SI changed depending on the study subpopulation (results not shown). Because the Predicted probability was calculated using the following formula: Abbreviations: SI, Survival index; predicted probability, predicted probability of all-cause death. mean age of this study was more than 60 years, the quadratic function of BMI may have been approximated as a linear function of BMI. It was suggested that the appropriate model of the relationship between BMI and mortality may depend on the study population.
GNRI is composed of serum albumin level and BMI as continuous variables, and GNRI has been reported to reflect HD patients' mortality [18]. CS is composed of categorical variables of serum albumin level, BMI and inflammation, and is associated with HD patients' mortality [19]. However, the c-statistics of GNRI and CS were lower than those of single indices (serum albumin level and BMI) depending on the outcome. GNRI coefficients were originally based on the data from a population of veterans (99% of them were male, non-ESRD) [17,32]. This suggests that an index with multiple variables may not always be accurate for the prediction of the outcome depending on the study population. On the other hand, although CS was developed on the basis of the HD-patient population, the cutoff values in CS may have not been appropriate for this study population. Thus, inappropriate cutoff values for a study may decrease the accuracy of CS for the prediction of the outcomes. These findings suggest that careful attention is required in using a risk index in a study population different from the population in which the risk index was developed. In this study, because SI also had this problem, stratification analysis for DM and vintage were examined to confirm the populations to which SI can be applied. The accuracy of SI in predicting death was higher than those of other indices in these populations examined. However, the problem was not completely eliminated, more studies are necessary to validate the usefulness of SI for various populations, such as those in various countries, and comorbid conditions.
In this study, some of the risk factors for death were included in SI. It was considered that these risk factors should be given priority in the population of this study. However, for example, DM as a cause of ESRD was not included in SI. A previous US DOPPS (DOPPS phases 1 and 2) also showed that DM as a cause of ESRD is not a statistically significant risk factor for one-year all-cause death [2]. On the other hand, another DOPPS showed that DM is commonly associated with mortality up to 5 years in patients from the US, Europe and Japan [33]. A significant difference in the prevalence of DM is observed in patients from the US, Europe and Japan [33]. These findings suggest that a factor such as DM is not always a statistically This study has several limitations. First, because of the observational nature of this study, the results may be biased by unmeasured confounders. Second, we were unable to examine the patients with missing data in this study, which might have caused selection bias. Third, the DOPPS datasets did not include enough nutritional data for assessing malnutrition, comorbid conditions and medications. We were unable to evaluate the effect of the differences in nutrition, comorbid conditions and medications on the risk of death. Further studies are needed to evaluate the relationship between these factors and SI. Fourth, because the DOPPS datasets did not include enough data to satisfy the requirements of other previously reported risk indices, SI could not be compared with those indices [34][35][36][37][38][39]. Fifth, the primary outcome of this study  was all-cause death within one year; thus, we were unable to evaluate transplantation as an outcome. Sixth, the negative coefficient of serum phosphorus level in SI suggests a possibility that malnourished patients with low serum phosphorus levels may have a higher SI than malnourished patients with high serum phosphorus levels. However, because malnourished patients usually have low BMI and serum creatinine, albumin and total cholesterol levels, the errors caused by serum phosphorus levels may be minimized.

Conclusions
The associations among many risk factors for death in HD patients are complex. This study showed a possibility that the simultaneous evaluation of multiple risk factors using SI can accurately assess patients' prognosis and identify patients at increased risk of death.