We conducted this retrospective study based on Medical Information Mart for Intensive Care IV version 1.0 (MIMIC-IV v1.0) database. MIMIC-IV, a large and public database, contains comprehensive data of more than 60,000 patients admitted to the ICU at Beth Israel Deaconess Medical Center from 2008 to 2019 (16). One author (P.J.C) has completed the online training course of the National Institutes of Health and obtained access to the database (record ID: 41046393). The project was approved by the institutional review boards of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center.

The inclusion criteria included: (1) age ≥ 65; (2) length of stay in ICU ≥ 24 h. Patients with missing key data (height, weight or albumin) on the first day of admission were excluded from the study. For patients with multiple hospitalizations, we only used their first hospitalization. The GNRI was calculated with the following formula (10): GNRI = 1.489 × serum albumin (g/L) + 41.7 × present weight (kg)/ideal weight (kg). The ideal body weight was calculated according to the Lorentz equations (10): 0.75 × height (cm) – 62.5 for men and 0.60 × height (cm) – 40 for women. When present weight exceeded ideal weight, present weight/ideal weight was set to 1. Patients were stratified into four groups according to the GNRI values, namely, major risk (GNRI: <82), moderate risk (GNRI: 82 to <92), low risk (GNRI: 92 to ≤98), and no risk (GNRI: >98) (10).

The primary outcome was all-cause mortality in hospital. The secondary outcomes included ICU mortality, length of stay (LOS) in ICU and LOS in hospital.

The PostgreSQL 10.7 software and Structured Query Language were used to extract the baseline data within the first 24 h of ICU admission from the MIMIC-IV database. The following variables were collected, (1) demographic characteristics (age, gender, height, weight); (2) laboratory indicators (white blood cell (WBC) count, platelet count, hemoglobin, alanine transaminase (ALT), aspartate aminotransferase (AST) international normalized ratio (INR), serum creatinine (sCr), blood urea nitrogen (BUN), serum sodium, serum potassium, serum chloride, bicarbonate, anion gap and lactate); (3) comorbidities were identified according to International Classification of Diseases, 9th revised (ICD-9) and 10th revised (ICD-10) editions (chronic obstructive pulmonary disease (COPD), congestive heart failure (CHF), myocardial infarction (MI), chronic kidney disease (CKD), cirrhosis, cerebral infarction, malignancy and sepsis); (4) clinical severity scales (Sequential Organ Failure Assessment (SOFA) score and Simplified Acute Physiology Score II (SAPS II)); (5) treatment measures (renal replacement therapy (RRT) and mechanical ventilation (MV)).

Continuous variables are presented as the mean ± standard deviation (SD) for normal distribution and as the median and interquartile range (IQR) for skewed distribution. Normal distributions were confirmed by Shapiro–Wilk test. Continuous variables were compared by one-way ANOVA or Kruskal-Wallis H test, respectively. Categorical variables were compared using the χ²-test or Fisher exact test as appropriate.

Multivariable Cox proportional hazard models were used to examine the hazard ratios (HRs) and 95% confidence intervals (CIs) for associations between predefined GNRI groups and mortality. Model 1 was adjusted for age, gender and laboratory indicators. Model 2 was adjusted for variables in model 1 plus comorbidities and treatment measures. Model 3 was adjusted for variables in model 2 plus clinical severities. The assumption of the proportional hazards analysis was confirmed graphically by log cumulative hazard plots for mortality based on GNRI category. p for trend test was conducted by including the levels of GNRI as an ordinal score to the regression models. Restricted cubic spline (RCS) is a powerful tool to characterize a dose–response association between a continuous exposure and an outcome. RCS divide the observed range of the continuous variable with k knots and create a third order polynomial above the knot. RCS fit smoothly at each knot and to be linear both below the first knot and above the last knot. The knots are usually located at fixed percentiles of the continuous variable (17). So, the associations between continuous scale of GNRI and mortality were evaluated by RCS based on Cox proportional hazard models with three knots at the 10^th, 50^th and 90^th percentiles of the distribution, adjusting for covariates in Model 3. The functional form of associations was evaluated by a Wald test comparing a linear or nonlinear likelihood ratio. Results were reported in log-relative hazard ratios and associated 95% CIs. If there were nonlinearity, we would further apply a two-piecewise linear regression model to examine the inflection point of GNRI on mortality, which provided maximum model likelihood. Finally, propensity score matching (PSM) was used to reduce selection bias in observational studies (18). Patients were matched in a 1:1 ratio with a caliper of 0.1 standard deviations of the Cox of the estimated propensity score with hospital mortality. Confounding factors such as age, gender, laboratory indicators, comorbidities, clinical severity scales and treatment measures were selected for matching.

Multivariable linear regression was used to analyze the relationship between GNRI (both as continuous and categorical variables) with LOS in hospital and ICU. Subgroup analyses according to gender, COPD, CHF, MI, CKD, cirrhosis, cerebral infarction, malignancy, sepsis (as defined by the Sepsis-3 criteria (19)), RRT and MV were conducted to test their interactions with GNRI on primary endpoint. GNRI was standardized to a Z-score ((GNRI- mean value)/SD) in order to present the confidence intervals of each subgroup clearly.

All data analyses were performed using R software (version 4.2.0; R Foundation for Statistical Computing, Vienna, Austria) and a two-sided p-value < 0.05 was considered statistically significant for all analyses. Variables with missing values were imputed using the multiple imputation method.

According to the inclusion criteria, a total of 3,696 elderly patients were finally obtained in the study (Figure 1). The median age of enrolled patients was 75 (IQR, 69–81) years with 2,095 (55.9%) male patients. Based on GNRI stratification, 1058 (28.6%) patients were in major risk group, 1180 (31.9%) patients were in moderate risk group, 743 (20.1%) patients were in mild risk group and 715 (19.3%) patients were in no risk group. The baseline characteristics of study population stratified by GNRI were shown in Table 1. With the decreasing of GNRI, patients with nutritional risk in ICU tended to be older and more likely to be female. In terms of laboratory indicators, patients in major risk group featured higher levels of WBC count, ALT, AST, bilirubin, INR, creatinine, BUN, and lactate and lower levels of hemoglobin, platelet count and bicarbonate compared with patients in no risk group. The prevalence of cirrhosis, sepsis and malignancy were more common in patients in major risk group. RRT and MV were used more frequently in patients with major nutritional risk. Clinical severities increased significantly with the decreasing of GNRI. When compared with patients in no risk group, patients in major risk group had significantly higher hospital mortality (30.7 vs. 15.2%, p < 0.001) and ICU mortality (24.3 vs. 11.7%, p < 0.001), and longer LOS in ICU (4.3, IQR (2.3–8.6) vs. 3.4, IQR (2.0–6.4), p < 0.001) and LOS in hospital [10.0, IQR (6.0–17.0) vs. 7.0, IQR (5.0–13.0), p < 0.001].

As shown in model 1, after adjusting for age, gender and laboratory indicators, multivariable Cox proportional hazard models demonstrated significant negative associations between GNRI categories and hospital mortality (major risk vs. moderate risk [HR = 0.79, 95% CI: 0.65–0.96]; vs. mild risk [HR = 0.74, 95%: CI 0.58–0.96]; vs. no risk [HR = 0.66, 95% CI: 0.49–0.88]; p for trend 0.002). In model 2, after adjusting for variables in model 1 plus comorbidities and treatment measures, GNRI categories still had significantly negative associations with all-cause hospital mortality (major risk vs. moderate risk [HR = 0.79, 95% CI: 0.65–0.96]; vs. mild risk [HR = 0.75, 95% CI: 0.58–0.97]; vs. no risk [HR = 0.66, 95% CI: 0.50–0.89]; p for trend 0.026). However, in model 3, after adjusting for variables in model 2 plus clinical severities, GNRI categories only featured negative trend with hospital mortality (p for trend 0.037). Besides, there were no significant correlations between GNRI categories and ICU mortality in model 3 (Table 2).

On a continuous scale of GNRI, restricted cubic spline in a fully adjusted model showed that the associations of GNRI levels with all-cause hospital mortality (p for non-linearity = 0.003) and ICU mortality (p for non-linearity = 0.032) were L-shaped (Figure 2). The two-piecewise linear regression models indicated that the inflection points of GNRI for hospital and ICU mortality were 78.7 and 78.9, respectively. At the left side of inflection point, GNRI levels were significantly negatively associated with hospital mortality (HR = 0.96, 95% CI: 0.94–0.98; p < 0.001) and ICU mortality (HR = 0.97, 95% CI: 0.94–0.99; p = 0.006). While, at the right side of inflection point, GNRI levels had no significant relations with hospital or ICU mortality (Table 3).

After PSM, 778 patients in the non-survivor group were matched with 778 patients in the survivor group. The baseline profiles were well balanced between the two groups with standardized mean differences <10% for most of the variables (Additional file 1: Supplementary Table 1). Restricted cubic spline in a fully adjusted model also revealed an “L-shaped” relation (p for non-linearity = 0.004) between GNRI and hospital mortality (Additional file 1: Supplementary Figure 1).

To further investigate possible interactions between GNRI levels and hospital mortality, several subgroup analyses were conducted according to gender, COPD, CHF, MI, CKD, cirrhosis, cerebral infarction, malignancy, sepsis, RRT and MV (Figure 3). After Z-transform standardization, significant interactions were observed in the subgroups of COPD (p for interaction =0.015) and malignancy (p for interaction =0.005). Elderly patients with COPD (HR = 0.62, 95% CI: 0.46–0.83; per unit increase in Z-score) and malignancy (HR = 0.74, 95% CI: 0.63–0.86; per unit increase in Z-score) featured stronger associations between GNRI levels and hospital mortality.

As both category and continuous variables, multiple linear regression showed that GNRI was negatively associated with length of stay in hospital even after adjustment for age, gender and clinical severities among patients who were survival. However, after adjustment for age, gender and clinical severities, there was no significant relation between GNRI and length of stay in ICU (Table 4).