In the present study, we retrospectively evaluated data from the Korean Society of Nephrology, which collects data regarding dialysis centers and patients through an online registry (http://www.ksn.or.kr). Details for the registry have been previously reported [16], and 763 centers (approximately 95% of all Korean dialysis units) participated in the registry during 2015. Baseline and retention data are recorded for patients who start dialysis, and in the present study we evaluated data from patients who started dialysis in 2000 (Fig. 1). Among 16,275 patients who started maintenance hemodialysis, 206 were excluded because they were < 18 years old (n = 72), subsequently underwent kidney transplantation (n = 131), or had incorrectly recorded dates of death (vs. registry data, n = 5). Thus, 16,069 patients were included in the analysis. Data regarding BMI were available for 10,327 patients, which we used in a complete case analysis, and multiple imputation of BMI was used for the 5,742 patients with missing BMI data. Complete data were available for all other included covariates (age, sex, and cause of ESRD). Missing patterns were tested for the missing completely at random assumption by Little’s test [17], which was rejected with P < 0.001. Dependence of BMI missing (missing at random assumption) was assumed by assessing whether missing indicators of BMI are dependent on other complete covariates. Uni- and multivariate logistic regression of BMI missing indicators on other complete covariates (age, sex, and cause of ESRD) were conducted, which showed dependency on age (P < 0.001) and cause of ESRD (P = 0.001 for hypertensive nephropathy, P < 0.001 for chronic glomerulonephritis, and P = 0.026 for other causes; P values from multivariable logistic regression).

Data for death status and date were acquired by linking the Korean Society of Nephrology hemodialysis dataset to Statistics Korea (http://kostat.go.kr). Patients who did not have records of death until December 31, 2013 were considered alive at that time with right censoring. When the hemodialysis registry data were linked to death status data from Statistics Korea, 5 cases could not be identified, which agrees with the number of patients with incorrectly recorded dates of death. After excluding these patients, there were no redundant matches. The data matching process was performed under the guidelines of strengthened privacy legislation in Korea, and the study protocol was approved by the public institutional review board of the Ministry of Health and Welfare (P01-201412-RS-02-02). Patient privacy was preserved in all instances, and the study methods complied with the tenets of the Declaration of Helsinki.

The baseline data for complete cases were analyzed according to the World Health Organization’s BMI categories [9], and inter-group differences were tested using the chi-square test or analysis of variance, as appropriate. For cases with missing BMI data, we used multiple imputation methods to estimate regression parameters. To briefly introduce multiple estimations, 30 randomly imputed datasets were generated by using mi commands in Stata. BMI missing values were imputed using other complete covariates. Regression analyses were conducted repeatedly per each dataset, and then all estimates were pooled to derive pooled estimated parameters with standard errors [18]. Time-to-event outcomes were evaluated by using the log-rank test with Kaplan–Meier estimator or Cox regression. The group-specific effects of BMI on 5-year mortality were evaluated using univariate or multivariate Cox regression analyses that were stratified using various subgroups. The age-dependent effects of BMI were estimated using three models: model 0 was unadjusted, model 1 was adjusted for only diabetes status and sex, and model 2 was adjusted for all included variables (cause of ESRD, age, and sex). To estimate the time-dependent hazard ratios for mortality, we applied time-varying effect models that were based on fractional polynomials [19–22]. All statistical analyses were performed using Stata software (version 15; StataCorp LLC, College Station, TX, USA). Data visualization was performed using Stata software and GraphPad Prism (version 6; GraphPad Software, LLC, San Diego, CA, USA).

Table 1 shows the baseline characteristics of the 10,327 patients who started hemodialysis during the study period. The mean patient age was 53.2 ± 13.9 years, and 5,914 patients (57.3%) were male. The main cause of ESRD was diabetes (4,525 patients, 43.8%), followed by hypertension (17.6%) and glomerulonephritis (13.7%). Mean BMI among all patients was 21.4 ± 3.1 kg/m², and most patients had normal BMI (18.5–24.9 kg/m²; n = 7,578, 73.4%). There were 1,563 patients (15.1%) who were underweight (< 18.5 kg/m²), and 126 patients (1.2%) who were obese (> 30 kg/m²). Obese patients had a lower mean age compared to the total patient population (48.0 ± 14.1 years vs. 53.2 ± 13.9 years, P = 0.001). Female sex was more common in the underweight group (53.2%) and in the obese group (60.3%). The obese group had a lower prevalence of diabetes-related ESRD, compared to the total patient population (39.7% vs. 43.8%). Hypertension was more common in the overweight and obese groups, compared to the total patient population (20.8% vs. 21.4% vs. 17.6%).

Fig. 2 shows the patient survivals of complete cases according to BMI category. During a median follow-up of 8.6 years (interquartile range, 3.3–11.2 years), 5,420 patients (52.5%) died. The survival rates of the BMI categories (log-rank P < 0.001) were significantly different, with underweight patients having the poorest survival and obese patients having the best survival throughout the follow-up period. For the regression analysis, we used 30 randomly imputed datasets and pooled estimates of individual datasets. There were no missing values for death or date of death, and among 16,069 patients, 9,272 patients (57.7%) died. Compared to the normal weight group, the underweight group had a significantly higher risk of mortality (hazard ratio [HR], 1.262; 95% confidence interval [CI], 1.179–1.350; P < 0.001), and this increased risk persisted in the various subgroups, regardless of age, sex, or diabetes status (Fig. 3). The overweight group had a reduced risk of mortality (HR, 0.904; 95% CI, 0.826–0.991; P = 0.031), and this risk was further reduced among overweight male patients (HR, 0.887; 95% CI, 0.791–0.996; P = 0.042), overweight diabetic patients (HR, 0.865; 95% CI, 0.776–0.964; P = 0.009), and overweight patients who were < 65 years old (HR, 0.882; 95% CI, 0.789–0.987; P = 0.028). However, the reduced risk of mortality disappeared among overweight female patients, overweight elderly patients, and overweight non-diabetic patients. The obese group had an even lower risk of mortality (HR, 0.589; 95% CI, 0.444–0.781; P < 0.001), and this reduced risk was observed in all subgroups except the elderly subgroup (> 65 years old) (Fig. 3). Table 2 shows the results of the multivariate Cox regression analyses, which were adjusted for age, sex, and cause of ESRD (including diabetic nephropathy). Underweight status was associated with a mortality HR of 1.292 (95% CI, 1.203–1.387) and overweight status was associated with a HR of 0.904 (95% CI, 0.829–0.985). Mortality was independently associated with age > 65 years (HR, 2.688; 95% CI, 2.571–2.811) and diabetes (HR, 1.845; 95% CI, 1.706–1.995). Female sex was inversely and independently associated with mortality (HR, 0.856; 95% CI, 0.821–0.893).

Table 3 presents the age-stratified multivariate analysis of BMI and mortality, as well as the crude and adjusted models. Underweight status was associated with poor survival across all age strata (young age: HR 1.306, 95% CI 1.011–1.687; middle age: HR 1.415, 95% CI 1.262–1.586; elderly: HR 1.239, 95% CI 1.130–1.358). The effects of different BMI groups were statistically significant in the middle age group, although overweight or obese statuses were not independently associated with mortality in the young or elderly groups. Fig. 4 presents the age-specific adjusted mortality HRs, compared to the young non-diabetic group. Aging was associated with an increased risk of mortality, although the underweight group had reductions in long-term overall mortality HR (Fig. 4B) compared to the 5-year mortality HR (Fig. 4A). Stepwise reductions in overall mortality were not observed with increasing BMI among the elderly subgroup, although this phenomenon was observed for overall mortality in the middle-age group. Obese young and middle-aged patients had slightly elevated changes in their overall mortality HR, which we further investigated using time-varying analyses.

In Fig. 2, underweight patients experienced the fastest decline in survival during the early period after starting hemodialysis, and also experienced the poorest survival during the follow-up period. The obese group experienced better survival until the median follow-up time, compared to the normal BMI group. This time-dependent change in the HRs was confirmed by testing a non-zero slope in a generalized linear regression of the scaled Schoenfeld residuals for time (Supplementary Fig. 1). Fig. 5 and 6 show the time-dependent mortality HRs according to BMI. Underweight status was associated with elevated HRs during the first 7 years after starting hemodialysis, while the obese group had decreased HRs during years 2 to 7 after starting hemodialysis. These time-varying patterns were similar for both sexes (Fig. 5). The young obese group experienced an increased risk of mortality beginning at 8 years after starting hemodialysis (Fig. 6), although this population was not large.