No incremental discriminative value of hemoglobin beyond GNRI for predicting mortality in older adult with sepsis

Abstract

Background

The prognostic value of hemoglobin-combined Geriatric Nutritional Risk Index (H-GNRI) in older adult septic patients remain unexplored. This study aimed to investigate the association between H-GNRI and both short-term and long-term mortality in older adult patients with sepsis.

Methods

This retrospective cohort study analyzed older adult with sepsis from the MIMIC-IV database. Patients aged ≥ 65 years meeting sepsis-3 criteria were included, excluding those with chronic kidney disease or severe liver disease. The H-GNRI scoring system was developed by combining optimal cutoff values for GNRI and hemoglobin determined through X-tile analysis. Propensity score matching (1:1) was performed to minimize confounding. The primary outcome was 28-day mortality. Cox proportional hazards models and ROC analysis were used to evaluate prognostic performance and discriminatory ability.

Results

The 28-day mortality rates increased significantly across risk groups (45.0%, 61.7%, and 72.5% for low, intermediate, and high-risk groups, respectively; P < 0.001). This trend persisted for 180-day and 1-year mortality. After adjusting for confounders, higher-risk H-GNRI groups remained independently associated with increased mortality risk. However, H-GNRI demonstrated poor discriminatory performance with an AUC of 0.58 for 28-day mortality prediction.

Conclusion

While H-GNRI demonstrated significant associations with mortality in older adult sepsis patients, it exhibited poor discriminatory performance, limiting its utility as a standalone risk prediction tool. Further studies are needed to develop more sophisticated nutritional risk stratification approaches for critically ill patients.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12877-026-07262-8.

Introduction

Sepsis, a life-threatening organ dysfunction caused by dysregulated host response to infection, resulting in a mortality rate of 20–40% [1, 2]. Critical illness triggers substantial metabolic and endocrine disturbances characterized by increased catabolism [3], insulin resistance [4], and altered nutrient utilization [5]. While these adaptive changes initially support cellular energy supply and host defense mechanisms, prolonged metabolic derangement can lead to malnutrition and other adverse outcomes [6, 7].

Malnutrition is a common complication in septic patients and is strongly associated with poor prognosis [8–11]. Current guidelines from the European Society for Clinical Nutrition and Metabolism (ESPEN) and the American Society for Parenteral and Enteral Nutrition (ASPEN) emphasize the importance of early nutritional support in critically ill patients [12], recommending initiation of enteral nutrition within 24–48 h of ICU admission. However, implementing early nutritional support in septic patients requires careful consideration and individualization. During the acute phase of sepsis, nutritional support must often be subordinate to hemodynamic resuscitation protocols, as premature enteral feeding may lead to complications ranging from emesis to intestinal ischemia. Notably, many clinical trials supporting early nutritional intervention have either excluded malnourished patients or included them in very small proportions, highlighting a significant gap in evidence-based guidance for this high-risk population 1. Therefore, comprehensive nutritional risk assessment becomes crucial for developing individualized nutritional support strategies.

The Geriatric Nutritional Risk Index (GNRI) [13], calculated using serum albumin and body weight, has emerged as a valuable tool for nutritional risk assessment and outcome prediction in various diseases [14, 15]. However, GNRI does not account for hemoglobin levels, which serve as both a nutritional indicator and an antimicrobial factor through inhibiting bacterial growth by limiting bacterial nutrient acquisition. In septic patients, hemoglobin levels are frequently decreased due to inflammatory response, infection-related consumption, and potential gastrointestinal bleeding. Previous studies have demonstrated that anemia is an independent risk factor for mortality in critically ill patients, with decreased hemoglobin levels significantly associated with organ dysfunction and increased mortality. The integration of hemoglobin into GNRI (H-GNRI) has been proposed and evaluated in various conditions [16–18]. However, a fundamental question remains: does combining multiple biomarkers truly enhance predictive performance compared to using individual indicators alone? While complex composite scores may seem intuitively superior, they risk increasing clinical complexity without proportional benefit. Therefore, this study aimed to systematically evaluate whether H-GNRI provides superior prognostic value compared to GNRI alone, and to determine the optimal prediction strategy for older adult patients with sepsis.

Material & methods

Data source

This retrospective study utilized data from a freely accessible critical care databases: the Medical Information Mart for Intensive Care IV (MIMIC-IV version 3.1) [19]. The use of this database was approved by the Institutional Review Board of the Massachusetts Institute of Technology. Authors Dr. Han Chen and Dr. Xiao-Yan Ding completed the required training and obtained authorization to access the database (certification numbers: HC 53297811 and XYD 55860595, respectively).

Study population

Inclusion criteria:

(1) Patients diagnosed with sepsis according to Sepsis-3 criteria; (2) Age ≥ 65 years; (3) First ICU admission; (4) ICU length of stay > 24 h.

Exclusion criteria (applied to both cohorts):

(1) Patients with chronic kidney disease (CKD) defined by ICD-9 codes (5851–5855, 5859) and ICD-10 codes (N18, N181-185, N189, 40301, 40311, 40391, 40402, 40403, 404112, 40413, 40492, 40493, E0822, E0922, E1022, EI122, EI322, I120, I311, I132); (2) Patients with liver cirrhosis or severe liver disease, where liver cirrhosis was defined by ICD-9 codes (5712, 5715–5719) and ICD-10 codes (K703, K740-K746, K7469), and severe liver disease was derived from the Charlson comorbidity); (3) Patients with missing data for hemoglobin, albumin, height, or weight measurements. The selection flowcharts are shown in Fig. 1 (MIMIC-IV).

Fig. 1.

Flow diagram of patient selection in older adult patients with sepsis

Data collection

We extracted demographic characteristics (age, gender, weight, height), laboratory measurements within 24 h of ICU admission (albumin level, hemoglobin level, bicarbonate, blood urea nitrogen (BUN), chloride, creatinine, platelets etc.), comorbidities (myocardial infarction, congestive heart failure, dementia, malignant cancer, chronic pulmonary disease, CKD, peptic ulcer disease, diabetes), and clinical parameters including vital signs (heart rate, systolic blood pressure (SBP), diastolic blood pressure (DBP), mean arterial blood pressure (MAP) and Sequential Organ Failure Assessment (SOFA) scores.

Missing data

To ensure the integrity of the data, the mean of this variable was replaced for variables with less than 5% missing values (bicarbonate, BUN, chloride, creatinine). For variables with outliers, use the WINSORIZE method, apply the WINSOR 2 command, with 1% and 99% as the cut-off points (height, weight, etc.). Missing data and abnormal data were processed using the STATA 17 software (Supplementary Table S1).

Exposure (H-GNRI assessment)

The GNRI was calculated as: [1.489 × albumin (g/L)] + [41.7 × (actual weight/ideal weight)] [14]. Ideal body weight was calculated using standard equations, with an adjustment factor for obesity (BMI > 30 kg/m²) [20]. The optimal cutoff values for hemoglobin and GNRI were determined using X-tile software [21], which identifies the cutoff points yielding the maximum chi-square value (corresponding to the minimum P-value) in the log-rank test. Patients were divided into three risk groups according to the H-GNRI score: low-risk, intermediate-risk and high-risk.

Outcomes

The primary outcome was 28-day all-cause mortality, with secondary outcomes including 180-day mortality, 1-year mortality, and length of ICU and hospital stay.

Propensity score matching (PSM)

PSM was performed to minimize potential selection bias and confounding factors between survival and non-survival groups. The selection of covariates for PSM was based on clinical relevance and variables that demonstrated significant differences between groups in the univariate analysis (P < 0.05). Covariates included demographic characteristics, vital signs, laboratory parameters, disease severity scores, and comorbidities.

A 1:1 nearest neighbor matching algorithm was employed with a caliper width of 0.1 standard deviations of the logit of the propensity score to ensure good matches while maintaining adequate sample size. The matching was performed without replacement, meaning that each control subject could be matched to only one treated subject. Balance between groups was assessed using standardized mean differences (SMD) before and after matching. An SMD < 0.1 was considered to indicate good balance between groups. The balance of covariates was visualized using love plots, which display the SMDs for each covariate before and after matching.

Statistical analysis

Continuous variables were presented as means ± standard deviation or medians (interquartile range) depending on their distribution, and categorical variables were expressed as numbers and percentages. Between-group comparisons were conducted using one-way analysis of variance (ANOVA) or Kruskal-Wallis test for continuous variables, and chi-square test or Fisher’s exact test for categorical variables as appropriate.

To address the dose-response relationship between GNRI/H-GNRI and mortality outcomes as suggested by previous literature, we employed two complementary analytical approaches: (1) For continuous GNRI values: Locally weighted scatterplot smoothing (LOWESS) regression was performed to visualize the non-parametric relationship between GNRI as a continuous variable and mortality rates at different time points (28-day, 180-day, and 1-year). (2) For categorical H-GNRI scores: Bar charts with overlaid trend lines were constructed to illustrate mortality rates across the three H-GNRI risk categories (high-risk, intermediate-risk, low-risk). One-way ANOVA was used to test for significant differences in mortality rates between groups.

Survival Analysis: Survival analysis was performed using Cox proportional hazards models, with hazard ratios (HRs) and 95% confidence intervals (CIs) reported. Variables were selected for the adjusted models based on both clinical relevance and statistical significance in univariate analysis (p < 0.05). The fully adjusted model (Model 4) included demographic factors (age, gender), disease severity (SOFA score), laboratory parameters (chloride, creatinine), and comorbidities (myocardial infarction, congestive heart failure, malignant cancer, and dementia). Kaplan-Meier curves were constructed to visualize survival differences between H-GNRI risk groups (low-risk, intermediate-risk, high-risk), with log-rank tests for comparison.

Predictive performance analysis

Areas under the ROC curves (AUCs) were calculated with 95% confidence intervals. Pairwise comparisons between AUCs were performed using DeLong tests to determine statistical significance of differences in predictive performance. All analyses were adjusted for potential confounders including age, sex, SOFA score, and relevant comorbidities. All statistical analyses were performed using Python (version 3.12.7) with scikit-learn, scipy, and pandas packages. Statistical significance was set at P < 0.05, and all tests were two-sided.

Sensitivity analysis

Given that X-tile analysis typically identifies multiple statistically significant cutoff values for mortality, sensitivity analyses were conducted to validate the robustness of our H-GNRI approach. While we selected the optimal combination yielding the highest chi-square value (hemoglobin 7.0 g/dL and GNRI 78.2), several alternative cutoff combinations also achieved statistical significance (p < 0.05) in the log-rank test. The discriminatory performance of these alternative H-GNRI combinations was evaluated using ROC analysis and compared with our optimal thresholds to assess whether the observed limited predictive value was attributable to suboptimal cutoff selection.

Results

As shown in Table 1, a total of 2270 patients initially included in the study, 1794 (79.0%) survived and 476 (21.0%) died. After PSM with a 1:1 ratio, 397 pairs of patients were successfully matched. After PSM, most variables achieved good balance with SMD < 0.1 (Fig. S1). The median age was similar between groups (78.6 vs. 78 years), and approximately half of the patients were female (52.6% vs. 50.6%). Disease severity was comparable, with both groups showing a median SOFA score of 7 points. Vital signs and hemodynamic parameters showed no significant differences, including systolic blood pressure (85 vs. 83 mmHg) and heart rate (109 vs. 110 bpm). Laboratory values, including renal function parameters and electrolytes, were also well-balanced between groups. The prevalence of major comorbidities was similar between survivors and non-survivors, including myocardial infarction (22.9% vs. 22.9%), congestive heart failure (38.3% vs. 40.3%), and diabetes (27.2% vs. 26.7%). Body mass index (BMI) was comparable between groups (26.34 vs. 26.79 kg/m²).

Table 1.

Baseline characteristics of elderly septic patients before and after propensity score matching

	Before PSM		SMD	After PSM		SMD
	Survival Group	Non Survival Group		Survival Group	Non Survival Group
	(n = 1794)	(n = 476)		(n = 397)	(n = 397)
Age (years)	75.3 [70.0, 81.9]	79.2 [71.7, 85.4]	0.341	78.6 [71.8, 85.6]	78.0 [71.5, 84.7]	0.052
BMI (kg/m²)	27.62 [24.07, 31.96]	26.22 [22.04, 30.71]	0.259	26.34 [23.00, 30.49]	26.79 [22.32, 31.14]	0.028
Gender (Female)	848 (47.3%)	252 (52.9%)	0.114	209 (52.6%)	201 (50.6%)	0.040
Bicarbonate (mmol/L)	21.00 [19.00, 24.00]	19.00 [16.00, 23.00]	0.370	20.00 [17.00, 23.00]	20.00 [16.00, 24.00]	0.004
BUN (mg/dL)	21.00 [15.00, 33.00]	31.00 [20.00, 51.00]	0.552	29.00 [19.00, 45.00]	30.00 [19.00, 47.00]	0.004
Chloride (mmol/L)	106.00 [103.00, 110.00]	105.00 [101.00, 109.00]	0.165	105.00 [101.00, 109.00]	105.00 [101.00, 109.00]	0.038
Creatinine (mg/dL)	1.10 [0.80, 1.40]	1.30 [1.00, 2.00]	0.291	1.20 [0.90, 1.80]	1.30 [0.90, 1.90]	0.069
Heart rate (bpm)	101 [88, 116]	114 [97, 131]	0.536	109 [93, 125]	110 [95, 126]	0.035
SBP (mmHg)	87 [79, 95]	81 [71, 91]	0.381	85 [76, 93]	83 [73, 93]	0.035
DBP (mmHg)	44 [39, 49]	42 [36, 49]	0.207	43 [38, 48]	43 [36, 49]	0.039
MAP (mmHg)	57 [52, 63]	55 [46, 62]	0.343	56 [49, 61]	56 [48, 63]	0.001
SOFA	5 [3, 7]	8 [5, 11]	0.811	7 [4, 10]	7 [5, 10]	0.047
Peptic ulcer disease (%)	59 (3.3%)	15 (3.2%)	0.008	17 (4.3%)	13 (3.3%)	0.053
Diabetes (%)	511 (28.5%)	125 (26.3%)	0.049	108 (27.2%)	106 (26.7%)	0.011
Chronic pulmonary disease (%)	471 (26.3%)	136 (28.6%)	0.052	118 (29.7%)	116 (29.2%)	0.011
Myocardial infarction (%)	321 (17.9%)	107 (22.5%)	0.117	91 (22.9%)	91 (22.9%)	0.000
Congestive heart failure (%)	559 (31.2%)	192 (40.3%)	0.196	152 (38.3%)	160 (40.3%)	0.041
Dementia (%)	126 (7.0%)	55 (11.6%)	0.168	52 (13.1%)	40 (10.1%)	0.094
Malignant cancer (%)	364 (20.3%)	179 (37.6%)	0.411	133 (33.5%)	136 (34.3%)	0.016
H-GNRI Risk
Low risk, (%)	1541 (85.9%)	308 (64.7%)		312 (78.6%)	269 (67.8%)	0.246
Intermediate risk, (%)	224 (12.5%)	136 (28.6%)	0.559	75 (18.9%)	105 (26.4%)
High risk, (%)	29 (1.6%)	32 (6.7%)		10 (2.5%)	23 (5.8%)

BMI Body mass index, BUN Blood urea nitrogen, DBP Diastolic blood pressure, GNRI Geriatric Nutritional Risk Index, H-GNRI Hemoglobin conbined Geriatric Nutritional Risk Index risk, MAP Mean arterial blood pressure, PSM Propensity score match, SBP Systolic blood pressure, SMD Sandardized mean difference, SOFA Sequential Organ Failure Assessment

The optimal cutoff values for hemoglobin and GNRI were determined using X-tile software (Supplementary, Figure S2). Multiple cutoff combinations achieved statistical significance (P < 0.05); however, the combination of hemoglobin 7.0 g/dL and GNRI 78.2 demonstrated the strongest discriminatory power with the largest effect size (P < 0.001) (Supplementary, Table S2). Based on these results, patients were divided into three risk groups according to the H-GNRI score: low-risk (H-GNRI score = 2, GNRI ≥ 78.2 and hemoglobin ≥ 7.0 g/dL), intermediate-risk (H-GNRI score = 1, either GNRI < 78.2 or hemoglobin < 7.0 g/dL), and high-risk (H-GNRI score = 0, both GNRI < 78.2 and hemoglobin < 7.0 g/dL). To address potential cutoff selection bias, we also evaluated a clinically relevant alternative combination (hemoglobin < 7.0 g/dL and GNRI < 90).

As shown in Table 2, the mortality demonstrated significant differences across H-GNRI risk groups. The 28-day mortality rates showed a clear ascending trend from low to high-risk groups (46.3%, 58.3%, and 69.7%, respectively; P = 0.001). This pattern persisted in longer-term outcomes, with 180-day mortality rates of 57.3%, 66.7%, and 81.8% (P = 0.002), and 1-year mortality rates of 62.5%, 72.8%, and 84.9% (P = 0.002) for low, intermediate, and high-risk groups, respectively.

Table 2.

Clinical outcomes across H-GNRI risk groups

	H-GNRI Risk
	Low Risk (n = 581)	Intermediate Risk (n = 180)	High Risk (n = 33)	P value
Length of ICU Stay, (days)	3.54 [1.93–6.56]	3.45 [2.08–6.81]	5.57 [2.46–10.02]	0.185
Length of Hospital Stay, (days)	8.64 [5.10-15.45]	10.93 [5.77–20.11]	17.96 [8.00-22.88]	0.001
28-day mortality, (%)	269 (46.3)	105 (58.3)	23 (69.7)	0.001
180-day mortality, (%)	333 (57.3)	120 (66.7)	27 (81.8)	0.002
1 year mortality, (%)	363 (62.5)	131 (72.8)	28 (84.9)	0.002

ICU intensive care unit, H-GNRI hemoglobin combined geriatric nutritional risk index

Length of hospital stay increased across risk groups, from 8.64 [5.10-15.45] days in the low-risk group to 10.93 [5.77–20.11] days in the intermediate-risk group, and 17.96 [8.00-22.88] days in the high-risk group (P = 0.010). Although a similar trend was observed in ICU length of stay (3.54 [1.93–6.56], 3.45 [2.08–6.81], and 5.57 [2.46–10.02] days, respectively), this difference did not reach statistical significance (P = 0.185).

Dose-response relationship between GNRI/H-GNRI and mortality

LOWESS regression analysis revealed a clear non-linear, inverse relationship between continuous GNRI values and mortality. As shown in Fig. 2 (panel A-C), survival probability increased progressively as GNRI values increased from 60 to 120, with the steepest improvement observed between GNRI values of 75–90. This consistent pattern across different follow-up periods supports a robust dose-response relationship.

Fig. 2.

Dose–Response Curves of GNRI and H-GNRI for 28-Day, 180-Day and 1-Year Mortality. A-C: LOWESS curves showing inverse relationship between GNRI values and mortality probability at 28-day (A), 180-day (B), and 1-year (C) timepoints. Mortality risk decreases most prominently as GNRI increases from 75 to 90. D-F: Mortality rates across H-GNRI categories (high-risk, intermediate-risk, low-risk) at 28-day (D), 180-day (E), and 1-year (F) timepoints, showing significant stepwise increases with declining risk scores (all p < 0.001)

As illustrated in Fig. 2 (panel D-F), survival rates demonstrated a significant stepwise increase from high-risk to low-risk H-GNRI categories at each follow-up interval (p < 0.001 for all comparisons). The 28-day survival rate increased from 17.7% in the high-risk group to 32.3% in the intermediate-risk group, and further to 62.0% in the low-risk group. Similar stepwise improvements were observed for 180-day survival and 1-year survival.

COX analysis results

Cox proportional hazards models were used to evaluate the association between H-GNRI risk categories and 28-day mortality. In the unadjusted model (Model 1), compared with the low-risk group, both high-risk and intermediate-risk groups showed significantly increased mortality risk (high-risk: HR 1.63, 95% CI 1.09–2.45, P = 0.019; intermediate-risk: HR 1.34, 95% CI 1.09–1.66, P = 0.006).

These associations remained consistent across different adjustment models. After controlling for age and gender (Model 2), SOFA score (Model 3), and further adjustment for laboratory parameters and comorbidities (Model 4), the high-risk group maintained significantly elevated mortality risk (HR 1.63, 95% CI 1.07–2.47, P = 0.021), as did the intermediate-risk group (HR 1.32, 95% CI 1.07–1.63, P = 0.011). Notably, none of the adjusted covariates showed significant independent associations with mortality in the fully adjusted model (Table 3).

Table 3.

Association between H-GNRI risk categories and 28-day mortality in older adult patients with sepsis: Univariate and multivariable cox regression analysis

	Univariate Analysis		Model 1		Model 2		Model 3		Model 4
	HR (95% CI)	P	HR (95% CI)	P	HR (95% CI)	P	HR (95% CI)	P	HR (95% CI)	P
H-GNRI Risk
High Risk	1.63 (1.09–2.45)	0.019	1.63 (1.09–2.45)	0.019	1.63 (1.08–2.46)	0.019	1.60 (1.06–2.42)	0.025	1.63 (1.07–2.47)	0.021
Intermediate Risk	1.34 (1.09–1.66)	0.006	1.34 (1.09–1.66)	0.006	1.34 (1.09–1.66)	0.006	1.33 (1.07–1.64)	0.009	1.32 (1.07–1.63)	0.011
Low Risk	1.53 (1.14–2.05)	0.004	reference	-	reference	-	reference	-	reference	-
Age	1.04 (1.02–1.05)	< 0.001			0.99 (0.90–1.08)	0.751	0.99 (0.90–1.08)	0.801	0.99 (0.90–1.09)	0.898
Gender	1.20 (1.00-1.45)	0.049			1.02 (0.85–1.23)	0.804	1.00 (0.85–1.19)	0.742	1.04 (0.87–1.25)	0.666
SOFA	1.20 (1.17–1.23)	< 0.001					1.02 (0.93–1.11)	0.421	1.02 (0.92–1.12)	0.704
Chloride	0.98 (0.96–0.99)	0.002							0.98 (0.89–1.07)	0.613
Creatinine	1.17 (1.11–1.24)	< 0.001							1.05 (0.96–1.16)	0.254
Myocardial Infarct	1.34 (1.07–1.67)	0.01							1.03 (0.83–1.28)	0.795
Congestive Heart Failure	1.43 (1.19–1.73)	< 0.001							1.04 (0.86–1.26)	0.703
Malignant Cancer	2.14 (1.77–2.59)	< 0.001							0.97 (0.80–1.18)	0.780
Dementia	1.53 (1.14–2.05)	0.004							0.85 (0.63–1.15)	0.293

Model 1: unadjusted; Model 2: adjusted for age and gender; Model 3: adjusted for Model 2 covariates plus SOFA score; Model 4: adjusted for Model 3 covariates plus laboratory parameters (chloride, creatinine) and comorbidities (myocardial infarction, congestive heart failure, malignant cancer and dementia). H-GNRI hemoglobin-geriatric nutritional risk index, SOFA sequential organ failure assessment

Similar patterns were observed in the analyses of 180 days mortality and 1-year mortality, with high-risk H-GNRI consistently associated with increased mortality risk (P < 0.01 for all analyses). None of the adjusted covariates showed significant independent associations with mortality in the fully adjusted model (Supplementary Table S3, S4).

The parallel model using alternative cutoffs (GNRI < 90 and hemoglobin < 7.0 g/dL) was additionally examined for comparison. Cox regression analyses showed no statistically significant associations between the risk groups and mortality endpoints (28-day, 180-day, or 1-year mortality) using the GNRI 90 cutoff (Supplementary Table S5-S7). These findings further confirm the superior discriminatory performance and prognostic validity of our data-driven cutoff (GNRI 78.2), as compared to the conventional cutoff of 90.

Kaplan-Meier analysis

Survival analysis demonstrated significant differences in survival probability among the three H-GNRI risk groups (log-rank P < 0.001). The study population included 581 patients (73.2%) in the low-risk group, 180 patients (22.7%) in the intermediate-risk group, and 33 patients (4.2%) in the high-risk group.

The survival curves showed early separation starting from the first few days of follow-up. At 28 days (Fig. 3A), distinct survival patterns were observed among the three groups, with the high-risk group showing the steepest decline in survival probability. This pattern persisted through 180 days (Fig. 3B) and remained consistent at 1-year follow-up (Fig. 3C). Throughout the entire observation period, the low-risk group maintained the highest survival probability, followed by the intermediate-risk group, while the high-risk group consistently demonstrated the lowest survival probability (log-rank P < 0.05 for all time points). While statistically significant, the survival curves showed considerable overlap, particularly between low and intermediate risk groups, indicating limited discriminatory ability.

Fig. 3.

Kaplan-Meier Survival Analysis of Short-term and Long-term Mortality According to Hb-GNRI Risk Stratification in older adult Patients with Sepsis. Kaplan-Meier curves showing (A) 28-day, (B) 180-day, and (C) 1-year survival probabilities stratified by Hb-GNRI risk categories in the propensity-matched cohort. Patients were classified into three risk groups: low risk (GNRI ≥ 78.2 and Hb ≥ 7.0 g/dL), intermediate risk (either GNRI < 78.2 or Hb < 7.0 g/dL), and high risk (both GNRI < 78.2 and Hb < 7.0 g/dL). The survival curves demonstrated significantly different mortality rates across risk groups (all log-rank P < 0.05). The number of patients at risk is shown below each graph. Shaded areas represent 95% CI. High-risk patients consistently showed worse mortality across all time points, while low-risk patients maintained the highest survival probabilities.

The parallel model using alternative cutoffs (GNRI < 90 and hemoglobin < 7.0 g/dL) also demonstrated significant survival differences in Kaplan-Meier analysis across all time points (28-day: P = 0.004; 180-day: P = 0.006; 1-year: P = 0.001), with comparable risk stratification patterns (Supplementary, Fig. S3).

ROC analysis

Qualitative approach: The H-GNRI scoring system (0–2 points) achieved an AUC of 0.583 (95% CI: 0.553–0.613), which was lower than GNRI alone [AUC 0.698 (95% CI: 0.670–0.726)] but higher than hemoglobin alone [AUC 0.558 (95% CI: 0.529–0.588)] (Fig. 4A). At optimal thresholds, sensitivity and specificity were 0.303 and 0.857 for H-GNRI, 0.739 and 0.537 for GNRI, and 0.476 and 0.618 for hemoglobin, respectively.

Fig. 4.

ROC Analysis for 28-Day Mortality Prediction. Panel A. Qualitative Approach ROC Analysis. The H-GNRI scoring system (0-2 points) achieved an AUC of 0.583 (95% CI: 0.553-0.613), which was lower than GNRI alone [AUC 0.698 (95% CI: 0.670-0.726)] but higher than hemoglobin alone [AUC 0.558 (95% CI: 0.529-0.588)]. Panel B. Quantitative Approach ROC Analysis. ROC curves comparing GNRI as a continuous variable with the combination model incorporating both GNRI and hemoglobin as continuous variables. GNRI maintained its discriminative performance (AUC 0.698, 95% CI: 0.670-0.726), while the combination model incorporating both GNRI and hemoglobin showed only marginal improvement (AUC 0.698, 95% CI: 0.670-0.727)

Quantitative approach: GNRI maintained its discriminative performance (AUC 0.698, 95% CI: 0.670–0.726), while the combination model incorporating both GNRI and hemoglobin showed only marginal improvement (AUC 0.698, 95% CI: 0.670–0.727) (Fig. 4B). At optimal cutoffs, the combination model achieved sensitivity and specificity of 0.587 and 0.697, respectively. Additional performance metrics for both approaches are detailed in Table 4 and 5.

Table 4.

ROC curve and predictive performance results for the qualitative approach

	AUC (95% CI)	Sensitivity	Specificity	PPV	NPV
GNRI	0.697 (0.669-0.726)	0.738	0.537	0.297	0.886
H-GNRI	0.583 (0.554-0.613)	0.304	0.857	0.36	0.823
Hemoglobin	0.558 (0.529-0.588)	0.228	0.866	0.310	0.809

Table 5.

ROC curve and predictive performance results for the quantitative modeling approach

	AUC (95% CI)	Sensitivity	Specificity	PPV	NPV
GNRI	0.698 (0.669-0.727)	0.738	0.537	0.297	0.886
Hemoglobin	0.558 (0.529-0.588)	0.228	0.866	0.31	0.809
GNRI+Hemoglobin	0.698 (0.669-0.726)	0.586	0.697	0.339	0.864

Sensitivity analysis

To assess whether the limited discriminatory performance was attributable to suboptimal cutoff selection, we evaluated all statistically significant H-GNRI combinations identified by X-tile analysis. Six alternative combinations were tested, including conventional GNRI thresholds (≥ 90) and other hemoglobin cutoffs: hemoglobin 7.4/GNRI 78.2, hemoglobin 9.0/GNRI 78.2, hemoglobin 7.0/GNRI 76.3, hemoglobin 7.4/GNRI 76.3, hemoglobin 7.4/GNRI 90, and hemoglobin 7.0/GNRI 90.

ROC analysis revealed that all alternative H-GNRI combinations demonstrated similar poor discriminatory performance for 28-day mortality prediction, with AUC values ranging from 0.52 to 0.58 (Fig. 5). No combination achieved clinically meaningful predictive accuracy (AUC > 0.70). Pairwise comparisons using DeLong tests showed no statistically significant differences between the optimal combination and alternative thresholds (all P > 0.05, Table S8-S9).

Fig. 5.

ROC Analysis for 28-Day Mortality Prediction. ROC curves for 28-day mortality prediction using different hemoglobin and GNRI threshold combinations. Each panel represents a different combination of cutoff values: A Hemoglobin 7.4 g/dL, GNRI 78.2; B Hemoglobin 9.0 g/dL, GNRI 78.2; C Hemoglobin 7.0 g/dL, GNRI 78.3; D Hemoglobin 7.4 g/dL, GNRI 78.3; E Hemoglobin 7.4 g/dL, GNRI 90; F Hemoglobin 7.0 g/dL, GNRI 90.

Discussion

In this cohort of older adult with sepsis, we found that adding hemoglobin to GNRI did not improve discriminatory performance for mortality prediction. Although H-GNRI stratified patients and showed statistically significant associations with mortality across follow-up periods in survival analyses, its predictive utility at the individual level was limited. The AUC for 28-day mortality was approximately 0.58, falling below commonly accepted thresholds for clinical usefulness, and there was no incremental value over GNRI alone. Even when hemoglobin and GNRI were combined as continuous variables in a logistic regression framework, ROC performance did not surpass GNRI. These findings contrast with reports from chronic disease settings where H-GNRI has outperformed GNRI [16–18], underscoring that results from stable conditions may not generalize to the acute pathophysiological milieu of sepsis [1, 22].

A plausible explanation for this discrepancy lies in the acute-phase perturbations characteristic of sepsis, which attenuate the signal content of traditional nutritional markers [3, 6]. Serum albumin, a core component of GNRI, functions both as a nutritional indicator and as an acute-phase reactant [23]. Inflammatory responses suppress hepatic synthesis and increase capillary permeability, while fluid resuscitation alters distribution volumes; together, these processes depress albumin levels and distort weight-related information within short time windows, lowering GNRI values irrespective of baseline nutritional status. Similarly, hemoglobin is strongly influenced by non-nutritional factors in sepsis, including hemodilution, cytokine-mediated suppression of erythropoiesis, reduced red cell survival, and iatrogenic blood loss [24, 25]. These dynamic, non-nutritional drivers dilute the “nutrition and oxygen-carrying” information that H-GNRI aims to capture, yielding statistical associations with outcome but insufficient discrimination for individual risk prediction.

Our sensitivity analyses support the robustness of these conclusions. To address potential information loss from categorical scoring, we evaluated the continuous combination of hemoglobin and GNRI; however, the AUC remained essentially unchanged compared with GNRI alone. Moreover, across multiple alternative cutoffs—including thresholds approximating those reported in prior studies —discrimination consistently remained suboptimal [16–18], indicating that the limited performance is not attributable to inappropriate cutoff selection but rather reflects intrinsic limitations of these markers under acute septic conditions. In direct comparisons, H-GNRI did not exceed GNRI and showed no clear advantage over hemoglobin alone, further weakening the rationale for the composite as a superior predictor in this population.

The cutoff findings warrant contextualization. The optimal GNRI cutoff identified in our septic cohort (78.2) was substantially lower than the conventional threshold (> 90) often cited for chronic conditions such as heart failure [16]. This observation is further supported by our dose-response analysis, where LOWESS curves revealed the most pronounced mortality reduction occurred when GNRI increased from very low values (< 80) to moderate values, aligning with our data-driven cutoff. When the traditional threshold was applied, Kaplan–Meier curves revealed group differences, yet the association did not persist in multivariable Cox models, suggesting insufficient independent prognostic value. By contrast, the 78.2 threshold yielded more robust associations in both univariable and multivariable analyses. This superior performance of our data-driven GNRI cutoff likely reflects the unique pathophysiological context of acute sepsis. This downward shift likely reflects acute inflammatory hypoalbuminemia rather than a redefinition of malnutrition in older adult with sepsis [26, 27]. While conventional GNRI thresholds were established in settings with relatively stable protein markers, the acute inflammatory response in sepsis rapidly alters these parameters, necessitating context-specific interpretative frameworks. Likewise, the hemoglobin threshold of 7.0 g/dL aligned with current transfusion guidelines [28], but incorporating hemoglobin into GNRI did not enhance discrimination, which is consistent with the predominance of non-nutritional determinants of hemoglobin in the acute phase [25, 29].

These observations have practical implications. In the unstable phase of sepsis, H-GNRI and GNRI appear to function more as background signals of illness severity and inflammatory burden than as high-precision tools for individualized risk prediction. As patients stabilize, these indices may still serve as pragmatic screening cues to prompt more comprehensive nutritional and functional assessment; however, their incremental contribution to immediate risk stratification is limited and should be interpreted with caution. Given the dynamic and heterogeneous nature of sepsis 1, future prognostic approaches should incorporate time-varying, sepsis-specific information—such as trajectories of inflammatory biomarkers, lactate and hemodynamic parameters, fluid balance, and organ function—followed by rigorous evaluation of calibration and decision-analytic utility rather than reliance on AUC alone.

The specific inclusion and exclusion criteria of this study warrant discussion regarding the physiological interpretation of GNRI. Severe liver and kidney diseases were excluded in our study to ensure the accuracy of nutritional assessment, as these conditions significantly disturb albumin metabolism and fluid balance (e.g., via impaired synthesis, ascites, or dialysis-dependent weight fluctuations), thereby compromising the validity of GNRI. Conversely, patients with heart failure were included in the final cohort. Given the high prevalence of heart failure in older septic patients, retaining this subgroup ensures the study’s results are generalizable to real-world clinical practice. However, it should be noted that, on the other hand, due to our strict exclusion of severe hepatic and renal diseases to ensure the internal validity of nutritional assessment, our findings regarding GNRI cannot be generalized to septic patients with concomitant cirrhosis or end-stage renal disease.

This study has limitations. First, the single-center, retrospective design constrains causal inference and generalizability. Second, the sample size was reduced after propensity score matching, which may have limited statistical power. Third, H-GNRI was calculated using admission values only, precluding assessment of dynamic changes during ICU stay—an important consideration in acute critical illness. Fourth, the poor discriminatory performance observed here may partly reflect the inherent noise of acute-phase markers; nonetheless, this characteristic is integral to the clinical reality of sepsis and reinforces the central conclusion that simple linear combinations of traditional nutritional indices have limited predictive value in this setting. Finally, as mentioned above, our inclusion and exclusion criteria may limit the generalizability to septic patients with concomitant cirrhosis or end-stage renal disease.

Conclusion

In summary, among older adult with sepsis, H-GNRI was statistically associated with mortality but exhibited poor discriminatory performance and provided no incremental value beyond GNRI. In this acute, dynamic context, pragmatic stratification may begin with GNRI using sepsis-appropriate thresholds, but individualized prediction will likely require dynamic models that integrate sepsis-specific biomarkers and repeated measurements, alongside thorough assessment of calibration and clinical decision value.

Abbreviations

ANOVA

Analysis of variance

ASPEN

American Society for Parenteral and Enteral Nutrition

BMI

Body mass index

BUN

Blood urea nitrogen

CI

Confidence interval

DBP

Diastolic blood pressure

ESPEN

European Society for Clinical Nutrition and Metabolism

GNRI

Geriatric Nutritional Risk Index

H-GNRI

Hemoglobin-combined Geriatric Nutritional Risk Index

HR

Hazard ratio

ICU

Intensive care unit

MAP

Mean arterial pressure

MIMIC-IV

Medical Information Mart for Intensive Care IV

MIT

Massachusetts Institute of Technology

NUTRIC

Nutrition Risk in the Critically ill

PSM

Propensity score matching

SBP

Systolic blood pressure

SMD

Standardized mean difference

SOFA

Sequential Organ Failure Assessment

Authors’ contributions

Han Chen contributed to the study conception and design, developed the methodology and software, conducted formal analysis, and secured funding for the project. Xiao-Yan Ding and Li-Juan Zeng were responsible for data curation, prepared the original draft of the manuscript, and contributed to funding acquisition. Jing-Ru Zhang and Xiao-E Zheng conducted the investigation and validation of research findings.

Funding

This work was supported by the Youth Top Talent Project of Fujian Provincial Foal Eagle Program; and the Startup Fund for scientific research, Fujian Medical University [grant number: 2021QH1290]; and the Startup Fund for scientific research, Fujian Medical University [grant number: 2022QH1312].

Data availability

The data used in this study were obtained from the MIMIC-IV, version 3.1, a publicly available critical care database. Access to the MIMIC-IV requires completed the training and obtained authorization.

Declarations

Ethics approval and consent to participate

The study utilized data from MIMIC-IV database. The use of the database was approved by the Institutional Review Board of Massachusetts Institute of Technology (MIT). The first author Xiao-Yan Ding and corresponding author Han Chen completed the required training and obtained authorization to access the database (certification numbers: XYD 55860595 and HC 53297811, respectively).

Consent to Publish

Not applicable.

Competing interests

The authors declare no competing interests.