Geriatric Nutritional Risk Index independently predicts mortality beyond frailty and sarcopenia screening in older adults: a retrospective cohort study

Abstract

Background

Malnutrition drives adverse outcomes in older adults, but routine screening is often hindered by time constraints and reliance on subjective data. The Geriatric Nutritional Risk Index (GNRI), derived from serum albumin and body weight, offers an objective alternative. However, its prognostic value relative to brief frailty and sarcopenia screens in outpatient settings remains unclear.

Methods

This retrospective cohort study included 1,288 patients aged ≥ 60 years from a tertiary geriatric outpatient clinic in Türkiye. The primary outcome was all-cause mortality. Multivariable Cox proportional hazards regression was used to estimate adjusted hazard ratios (aHRs), controlling for age, sex, and estimated glomerular filtration rate (eGFR). Discriminative performance and incremental prognostic value were compared against the FRAIL scale, SARC-F questionnaire, and Mini Nutritional Assessment–Short Form (MNA-SF) at a primary 3-year landmark.

Results

During a median follow-up of 581 days, 134 deaths (10.4%) occurred among 1,288 participants (median age 70 years; 67.5% female). Each one-unit increase in the GNRI was associated with an 8% lower mortality risk (aHR 0.92; 95% CI 0.90–0.95; p < 0.001). This association remained significant when the GNRI was evaluated simultaneously with the MNA-SF, FRAIL, and SARC-F. At the 3-year landmark (n = 367), the GNRI achieved the highest numerical discrimination among all screening tools (AUC 0.701; 95% CI 0.636–0.766), though pairwise differences did not reach statistical significance after Holm adjustment. Adding the GNRI to the base clinical model yielded a statistically significant incremental value (ΔC-statistic = 0.044; p < 0.001), representing the study’s most robust discriminative finding. Results were consistent across age, sex, and renal function subgroups.

Conclusions

In geriatric outpatients, the GNRI independently predicts all-cause mortality and provides significant incremental value beyond a base clinical model. While it showed the highest numerical discrimination among frailty and sarcopenia screens, this standalone advantage did not reach statistical significance, supporting its integration as a complementary tool in routine geriatric assessment. A data-driven threshold of ≤ 105 for screening and the established cutoff of ≤ 98 for rule-in support a tiered approach to nutritional risk stratification. Prospective, multicentre validation is warranted before widespread clinical adoption of these specific cutoffs.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12877-026-07597-2.

Background

Malnutrition is a major geriatric syndrome and a strong predictor of morbidity, institutionalisation, and mortality [1, 2]. It drives these adverse outcomes through interconnected pathways, including systemic inflammation (e.g., elevated interleukin-6, C-reactive protein, and tumour necrosis factor-α), impaired immunity, accelerated sarcopenia, and compromised wound healing. Overlapping with frailty, these mechanisms exacerbate reduced physiological reserve, functional decline, and vulnerability to stressors [3, 4]. Therefore, timely identification of high-risk individuals in outpatient care is essential for targeted intervention.

Although comprehensive nutritional and functional assessments are available, their routine implementation is hindered by time constraints and a reliance on patient-reported data. The FRAIL scale and SARC-F questionnaire rely entirely on subjective reporting, which may be unreliable in cognitively impaired populations [5, 6]. The Mini Nutritional Assessment–Short Form (MNA-SF) incorporates objective anthropometrics (body mass index and calf circumference) but depends heavily on recalled dietary intake and self-reported weight loss. While these tools are routinely used to evaluate physical decline and malnutrition risk, the extent to which they overlap with, or substitute for, entirely objective indices remains uncertain.

In contrast, the Geriatric Nutritional Risk Index (GNRI) is a fully objective screening tool derived exclusively from routine laboratory (serum albumin) and anthropometric (body weight and height) data [7, 8]. It requires no additional testing, patient cooperation, or specialised training, making it highly efficient for high-volume outpatient clinics where comprehensive assessments are impractical. Furthermore, by incorporating serum albumin, the GNRI captures the combined effects of nutritional depletion and systemic inflammation independently of patient recall [9, 10].

The GNRI is well-established in hospitalised and institutionalised populations [11–14]. However, its prognostic value remains poorly characterised in ambulatory settings, where nutritional depletion is less prevalent but early identification is highly impactful. To our knowledge, no study has directly compared the GNRI against frailty and sarcopenia screens for predicting mortality in an outpatient geriatric cohort. Clarifying whether this objective, laboratory-based index offers comparable or superior prognostic accuracy to established subjective screens is essential for optimising outpatient assessment workflows.

We hypothesised that the GNRI supplements, rather than replaces, brief functional screens by capturing biological markers of vulnerability that subjective questionnaires miss. Accordingly, this study aimed to evaluate the independent prognostic value of the GNRI for long-term all-cause mortality in a cohort of Turkish geriatric outpatients. By comparing its discriminative performance against the FRAIL scale, SARC-F, and MNA-SF, we sought to determine whether the GNRI provides an objective complement to existing tools, thereby streamlining comprehensive risk stratification.

Methods

Study design, setting, and participants

This study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. This single-centre, retrospective cohort study was conducted at an outpatient geriatric clinic in Türkiye. As a single-centre study, the findings should be interpreted within the context of the referral patterns and population characteristics of a tertiary geriatric outpatient clinic in central Türkiye. Consecutive adults aged ≥ 60 years who underwent a comprehensive geriatric assessment between April 2018 and May 2025 were screened. Patients were excluded if they: (i) presented with an acute condition requiring same-day hospitalisation or emergency department transfer, as determined by the attending geriatrician at the index visit based on clinical assessment; (ii) had metastatic cancer documented in the electronic medical record, because cancer-related cachexia confounds nutritional indices and imposes a competing mortality risk that would obscure the GNRI–mortality relationship under study; (iii) had missing data for GNRI calculation; or (iv) had a missing index visit date. Of 1,289 patients screened, one was excluded due to a missing index visit date, yielding a final analytic cohort of 1,288 participants.

Participants were followed from the index assessment until death or administrative censoring on 1 June 2025. The primary outcome was all-cause mortality. Vital status was ascertained through deterministic linkage with the National Death Notification System using Turkish Citizenship Numbers as unique identifiers. The registry is updated continuously for all citizens, ensuring complete and timely capture of mortality events.

Data collection and variable definitions

Baseline demographic, anthropometric, clinical, and laboratory variables were extracted from electronic medical records. Comorbidities were identified on the basis of physician-documented diagnoses and medication records. Chronic kidney disease (CKD) was defined as an estimated glomerular filtration rate (eGFR) < 60 mL/min/1.73 m² [15]. eGFR was calculated using the 2021 creatinine equation [16]. Anaemia was defined as haemoglobin < 13 g/dL in men and < 12 g/dL in women [17].

Geriatric and nutritional assessment

Functional and nutritional status were assessed at baseline using four validated tools:

Frailty (FRAIL scale)

The FRAIL scale was used due to its feasibility in routine clinical practice and validated performance in the Turkish geriatric population [18]. The scale includes five components (Fatigue, Resistance, Ambulation, Illnesses, Loss of weight) and yields a total score of 0–5. Scores of 3–5 indicate frailty.

Sarcopenia risk (SARC-F)

This questionnaire assesses Strength, Assistance with walking, Rise from a chair, Climb stairs, and Falls. A score of 4 or higher was considered indicative of probable sarcopenia [19].

Subjective nutritional status (MNA-SF)

The Mini Nutritional Assessment–Short Form (MNA-SF) was used to screen for malnutrition. Scores of 0–7 indicate malnutrition, 8–11 indicate risk of malnutrition, and 12–14 indicate normal nutritional status [20].

Objective nutritional status (GNRI)

The Geriatric Nutritional Risk Index was calculated using the original formula [7], where serum albumin measured in g/dL was converted to g/L by multiplying by 10:

Body weight was measured using a calibrated electronic scale with patients in light clothing without shoes. Height was measured using a wall-mounted stadiometer. Both measurements were obtained at the index outpatient visit.

Ideal body weight was calculated using the Lorentz formula:

If actual body weight exceeded ideal body weight, the weight ratio was set to 1. A GNRI score ≤ 98 defined nutritional risk [7]. Alternative GNRI thresholds were explored using time-dependent diagnostic accuracy criteria.

Statistical analysis

Descriptive statistics and group comparisons

Normality of continuous variables was assessed using the Shapiro–Wilk test. Continuous variables are presented as medians (interquartile ranges [IQR]) and categorical variables as n (%). Between-group comparisons used the Mann–Whitney U test for continuous variables and the chi-square or Fisher’s exact test for categorical variables.

Survival analysis

Time-to-event analyses used Kaplan–Meier survival curves with log-rank tests. Multivariable Cox proportional hazards regression identified independent predictors of all-cause mortality. The Efron approximation handled tied event times, and Wald-type 95% confidence intervals (CIs) were computed for all hazard ratio (HR) estimates. The primary multivariable model included age, sex, eGFR, and GNRI, selected a priori to maintain adequate events per variable (> 10). Post-hoc power analysis confirmed 99.8% power (α = 0.05, two-sided) to detect the observed GNRI hazard ratio (HR = 0.92 per unit) in the primary Cox model, using the Hsieh and Lavori formula accounting for covariate correlation (R² = 0.07). However, pairwise AUC comparisons between screening tools were underpowered (8–23%) to detect the observed differences of 0.027–0.050, which would require 1,880–7,527 total subjects at the 3-year landmark for 80% power.

Additional geriatric screening tools (MNA-SF, SARC-F, FRAIL) were evaluated in separate adjusted models and a combined model. Body mass index (BMI) and comorbidity count were tested as additions to the primary model using likelihood ratio tests. The proportional hazards assumption was evaluated using Schoenfeld residuals. Multicollinearity was assessed using variance inflation factors (VIF); all covariates had VIF < 1.4 (GNRI: 1.08, age: 1.33, eGFR: 1.28, sex: 1.05), indicating no meaningful collinearity between GNRI components and eGFR despite the biological correlation between serum albumin and renal function.

GNRI was analysed as: (1) a continuous variable (HR per unit); (2) a binary variable (≤ 98 vs. >98); and (3) a four-category variable (major risk: < 82; moderate risk: 82 to < 92; low risk: 92–98; no risk: > 98) [7]. The linearity of the association between GNRI and log-hazard was assessed using restricted cubic splines.

Discrimination and calibration

Model discrimination was quantified using Harrell’s C-statistic [21]. Receiver operating characteristic (ROC) analyses were performed for GNRI and comparator tools at 1-year, 3-year (primary), and 5-year horizons using a landmark binary approach. Sensitivity analyses using cumulative/dynamic ROC analysis with inverse probability of censoring weighting were performed to assess the influence of censoring [22]. To characterise potential selection bias from landmark exclusion, standardised mean differences (SMDs) between included and excluded subjects were computed (imbalance threshold: SMD > 0.2).

Pairwise comparisons of discriminative performance were conducted at the 3-year horizon using the DeLong method [23], with Holm-adjusted p-values for multiple testing. Diagnostic accuracy metrics (sensitivity, specificity, positive and negative likelihood ratios, and diagnostic odds ratio) were calculated at the Youden-optimal cutoff. Confidence intervals (95%) were derived using bias-corrected and accelerated bootstrap resampling (5,000 iterations).

Calibration was assessed using the Brier score and calibration plots generated via locally weighted scatterplot smoothing. Internal validation was performed using bootstrap resampling (1,000 iterations) to obtain optimism-corrected estimates. The Hosmer–Lemeshow goodness-of-fit test was applied to logistic models for the 3-year binary outcome.

Reclassification and clinical utility

Category-free net reclassification improvement (NRI) and integrated discrimination improvement were calculated from logistic regression models on the 3-year landmark sample (n = 367) [24]. Category-based NRI was computed using risk thresholds of 10% and 30%. Decision curve analysis was performed to assess clinical utility [25].

Subgroup and sensitivity analyses

Interactions between GNRI and age, sex, and eGFR (modelled continuously) were tested using likelihood ratio tests comparing models with and without the cross-product term. Subgroup analyses were performed by sex, age (< 75 vs. ≥ 75 years), and CKD status. Sensitivity analyses included re-estimating the primary model after excluding early deaths (< 30 days) and using multiple imputation (20 imputations) for models involving haemoglobin.

Statistical software and missing data

All tests were two-sided, and p < 0.05 was considered statistically significant. Analyses were performed using R version 4.5.2 (R Foundation for Statistical Computing, Vienna, Austria) with the survival, survminer, timeROC, pROC, boot, Hmisc, rms, and tableone packages.

The primary model covariates (age, sex, eGFR, GNRI) and outcome variables (mortality, follow-up) had no missing data (n = 1,288). BMI had minimal missingness (1.2%); complete-case analysis was used for models incorporating BMI. Haemoglobin, not included in the primary model, had 22.2% missingness. Little’s MCAR test rejected completely random missingness (χ² = 302.8, df = 56, p < 0.001); comparison of patient characteristics by haemoglobin availability indicated that missingness was associated with observed covariates (primarily renal function), supporting the Missing At Random (MAR) assumption. Sensitivity analyses incorporating haemoglobin employed multiple imputation (20 imputations, predictive mean matching) under the MAR assumption.

Results

Baseline characteristics and mortality

A total of 1,288 older adults were included (median age 70 years, IQR 66–75; 67.5% female). During a median follow-up of 581 days (1.6 years; range 4–2,400 days), 134 participants (10.4%) died.

Baseline characteristics stratified by GNRI status are presented in Table 1. Overall, 4.0% of participants (n = 51) had a GNRI ≤ 98. Given the small size of the at-risk subgroup, between-group absolute differences and 95% confidence intervals are provided in Supplementary Table S1. Compared with participants at no nutritional risk (> 98), the at-risk cohort was significantly older, predominantly male, and had lower BMI, serum albumin, and haemoglobin levels, along with a higher prevalence of CKD (Table 1). Geriatric syndrome burden was substantially higher, with significantly worse MNA-SF, SARC-F, and FRAIL scores (all p < 0.001). All-cause mortality was markedly higher among participants at nutritional risk (41.2% vs. 9.1%; absolute risk difference 32.0% points, 95% CI 19.5–45.8; Supplementary Table S1).

Table 1.

Baseline characteristics of the study population stratified by GNRI nutritional risk status

Characteristic	Overall (n = 1,288)	No Risk, GNRI > 98 (n = 1,237)	At Risk, GNRI ≤ 98 (n = 51)	p-value
Demographics
Age, years, median (IQR)	71.3 (66.0–75.2)	70.0 (66.0–75.0)	73.0 (69.5–77.5)	0.002
Female sex, n (%)	870 (67.5)	847 (68.5)	23 (45.1)	< 0.001
Anthropometrics
BMI, kg/m², median (IQR)	30.3 (26.6–34.9)	30.4 (26.8–35.0)	28.0 (22.4–33.1)	0.006
Laboratory values
Albumin, g/dL, median (IQR)	4.5 (4.2–4.7)	4.5 (4.3–4.7)	3.6 (3.3–3.7)	< 0.001
eGFR, mL/min/1.73 m², median (IQR)	77.5 (61.1–91.3)	77.6 (61.4–91.4)	66.2 (46.3–86.8)	0.015
Hemoglobin, g/dL, median (IQR) b	13.8 (12.7–14.7)	13.8 (12.8–14.8)	12.1 (10.2–13.4)	< 0.001
Comorbidities, n (%)
Chronic kidney disease	299 (23.2)	276 (22.3)	23 (45.1)	< 0.001
Diabetes mellitus	634 (49.2)	607 (49.1)	27 (52.9)	0.690
Hypertension	822 (63.8)	790 (63.9)	32 (62.7)	0.989
Coronary artery disease	260 (20.2)	245 (19.8)	15 (29.4)	0.094
Dementia	101 (7.8)	95 (7.7)	6 (11.8)	0.283
Geriatric screening tools
MNA-SF, median (IQR)	12.0 (10.0–13.0)	12.0 (10.0–13.0)	10.0 (7.5–12.0)	< 0.001
SARC-F, median (IQR)	3.0 (1.0–5.0)	2.0 (1.0–5.0)	5.0 (2.0–8.0)	< 0.001
FRAIL scale, median (IQR)	2.0 (1.0–3.0)	2.0 (1.0–3.0)	3.0 (2.0–4.0)	< 0.001
Outcome
All-cause mortality, n (%)	134 (10.4)	113 (9.1)	21 (41.2)	< 0.001

^a BMI available for n = 1,273 (15 missing, 1.2%). ^b Hemoglobin available for n = 1,002 (286 missing, 22.2%)

Abbreviations: BMI Body mass index, eGFR Estimated glomerular filtration rate, FRAIL Fatigue, Resistance, Ambulation, Illnesses, Loss of weight, GNRI Geriatric Nutritional Risk Index, IQR Interquartile range, MNA-SF Mini Nutritional Assessment–Short Form, SARC-F Strength, Assistance walking, Rising from chair, Climbing stairs, Falls

To assess the extent to which the GNRI and MNA-SF identify the same at-risk population, a cross-tabulation of GNRI risk status (≤ 98 vs. >98) against MNA-SF categories was performed (Supplementary Table S2). Among 51 patients with GNRI ≤ 98, two-thirds (n = 34; 66.7%) were also classified as malnourished or at risk of malnutrition by the MNA-SF (score ≤ 11), whereas 17 (33.3%) had normal MNA-SF scores (12–14), suggesting objective nutritional depletion not captured by the questionnaire. Conversely, of 625 patients classified as malnourished or at risk by the MNA-SF, only 34 (5.4%) had a GNRI ≤ 98, indicating that 94.6% of MNA-SF-positive patients would not be flagged by the standard GNRI threshold. Agreement between the two binary classifications was poor (Cohen’s κ = 0.030; p = 0.008). When mortality was examined across four concordance–discordance groups, both the concordant high-risk group (GNRI ≤ 98 and MNA-SF ≤ 11; 41.2%) and the GNRI-only group (GNRI ≤ 98 with normal MNA-SF; 41.2%) had markedly higher mortality than the MNA-SF-only group (12.5%) and the concordant low-risk group (6.0%; Fisher’s exact test, p < 0.001; Supplementary Table S2).

Survival analysis

Participants at nutritional risk (GNRI ≤ 98) demonstrated significantly lower cumulative survival (log-rank p < 0.001; Fig. 1). When stratified by the four-category classification, survival decreased progressively with worsening nutritional risk (log-rank p < 0.001; Additional file 6). Mortality rates increased from the no-risk group (9.1%) to the low-risk (38.9%) and moderate-risk (54.5%) groups, though the major-risk category contained too few patients (n = 4) to establish a reliable estimate.

Fig. 1.

Kaplan–Meier survival curves for all-cause mortality stratified by the Geriatric Nutritional Risk Index (GNRI): No Risk (GNRI > 98) versus At Risk (GNRI ≤ 98). The dashed lines indicate the median survival time for the At Risk group. Shaded areas represent 95% confidence intervals. Numbers at risk are shown below the panel at one-year intervals. Log-rank test p < 0.001

Multivariable predictors of mortality

In univariate analysis, an at-risk GNRI (≤ 98) was associated with a more than five-fold increase in mortality (HR 5.45; 95% CI 3.41–8.70; p < 0.001). Other significant univariate predictors included worse MNA-SF, SARC-F, and FRAIL scores, lower BMI and serum albumin, lower eGFR, recent weight loss, and CKD. Traditional cardiovascular comorbidities, dementia, depression, and recent falls were not significantly associated with mortality.

In the primary multivariable model (Table 2), older age, male sex, lower eGFR, and lower GNRI remained independently associated with mortality. A linear relationship was observed between continuous GNRI scores and mortality risk. Adjusted categorical models demonstrated a dose–response gradient, with mortality risk rising significantly in the low-risk (aHR 3.15; 95% CI 1.78–5.55) and moderate-risk (aHR 5.33; 95% CI 2.31–12.27) groups relative to patients at no risk.

Table 2.

Cox proportional hazards regression models for all-cause mortality

Variable	Univariate HR (95% CI)	p-value	Adjusteda HR (95% CI)	p-value	Adjusted + BMIb HR (95% CI)	p-value
Age, per year	1.08 (1.06–1.11)	< 0.001	1.05 (1.02–1.07)	< 0.001	1.05 (1.02–1.07)	< 0.001
Female sex	0.34 (0.24–0.48)	< 0.001	0.43 (0.30–0.61)	< 0.001	0.42 (0.29–0.62)	< 0.001
eGFR, per mL/min/1.73 m²	0.98 (0.98–0.99)	< 0.001	0.99 (0.98–1.00)	0.030	0.99 (0.98–1.00)	0.019
GNRI, per unit	0.90 (0.89–0.92)	< 0.001	0.92 (0.90–0.95)	< 0.001	0.93 (0.90–0.95)	< 0.001
BMI, per kg/m²	0.94 (0.91–0.97)	< 0.001	—	—	0.99 (0.96–1.03)	0.78
MNA-SF, per unitc	0.83 (0.78–0.89)	< 0.001	0.85 (0.80–0.91)	< 0.001	—	—
SARC-F, per unitc	1.18 (1.11–1.25)	< 0.001	1.19 (1.12–1.27)	< 0.001	—	—
FRAIL scale, per unitc	1.45 (1.28–1.65)	< 0.001	1.45 (1.27–1.64)	< 0.001	—	—

Primary adjusted model discrimination: C-statistic 0.75 (95% CI 0.70–0.80). Proportional hazards assumption met (global Schoenfeld test, p = 0.87). All variance inflation factors below 1.5

^a Adjusted for age, sex, and eGFR. ^b Additionally adjusted for BMI. BMI did not significantly improve model fit (likelihood ratio test, p = 0.72). Comorbidity count was also tested and did not improve fit (p = 0.94). ^c Each tool entered in a separate model adjusted for age, sex, and eGFR. When all tools were entered simultaneously, only GNRI (p < 0.001) and MNA-SF (p = 0.03) retained significance; SARC-F (p = 0.13) and FRAIL (p = 0.08) did not

Abbreviations: BMI Body mass index, CI Confidence interval, eGFR Estimated glomerular filtration rate, GNRI Geriatric Nutritional Risk Index, HR Hazard ratio, MNA-SF Mini Nutritional Assessment–Short Form, SARC-F Strength, Assistance with walking, Rising from a chair, Climbing stairs, and Falls

While MNA-SF, SARC-F, and FRAIL were each independently associated with mortality in separate adjusted models, simultaneous analysis of all four tools revealed that only GNRI and MNA-SF retained statistical significance (Table 2). Incorporating BMI or total comorbidity count into the primary model did not improve model fit (Table 2). The discriminative value of GNRI for predicting mortality was equivalent to serum albumin alone and superior to BMI alone (Additional file 1: Tables S3).

Diagnostic performance and screening tool comparison

GNRI maintained good discriminative ability across the 1- and 3-year horizons (Table 3; Additional file 4). SMDs between landmark-included and landmark-excluded subjects remained below the prespecified imbalance threshold at these horizons, though meaningful differences emerged at 5 years (Additional file 3). At the primary 3-year endpoint, GNRI achieved an AUC of 0.701 (95% CI 0.636–0.766). The 5-year estimates should be regarded as exploratory owing to the limited number of long-term survivors (n = 45).

Table 3.

Landmark binary ROC performance and optimal diagnostic cutoffs for GNRI

Time point	n	Events	AUC (95% CI)	Optimal cutoff (95% CI)	Sensitivity, % (95% CI)	Specificity, % (95% CI)
1-year	767	80	0.679 (0.604–0.745)	≤ 104.9	56.2 (44.9–67.1)	79.8 (76.4–82.6)
3-yeara	367	122	0.701 (0.636–0.766)	≤ 105.0	51.6 (42.7–60.7)	87.8 (83.1–91.4)
5-yearb	179	134	0.715 (0.633–0.785)	≤ 104.8	49.3 (40.6–57.7)	95.6 (84.5–100.0)
Clinical cutoffc	367	122	—	≤ 98	16.4 (10.4–23.7)	97.6 (94.8–99.2)

^a Primary endpoint. ^b Limited data: only 45 survivors at the 5-year landmark; wide confidence intervals reflect instability. ^c Established clinical cutoff per Bouillanne et al. [7]. All CIs derived using 5,000 bootstrap replicates. Extended diagnostic metrics in Additional file 1: Table S4

Abbreviations: AUC Area under the receiver operating characteristic curve, CI Confidence interval, GNRI Geriatric Nutritional Risk Index

At the 3-year endpoint, a Youden-optimal cutoff of ≤ 105.0 yielded a balanced diagnostic profile (Table 3). The established clinical cutoff of ≤ 98 achieved near-complete specificity at the cost of lower sensitivity. Analysis across graded GNRI risk strata demonstrated a clear dose–response relationship, with severe nutritional risk (≤ 92) yielding the highest positive likelihood ratio for mortality (Additional file 1: Table S4).

Among the four geriatric screening tools, GNRI demonstrated the highest nu merical standalone discrimination for 3-year mortality (AUC 0.701), followed by MNA-SF, FRAIL, and SARC-F (Table 4). While FRAIL achieved the highest sensitivity at its optimal cutoff, it had the lowest specificity; GNRI provided the most balanced trade-off between sensitivity and specificity (Table 4; Fig. 2). Pairwise AUC comparisons did not reach statistical significance after Holm adjustment (Table 5). However, adding GNRI to the base clinical model provided the largest incremental value of any single tool evaluated (Table 6).

Table 4.

Comparison of geriatric screening tools for predicting 3-year all-cause mortality

Tool	AUC (95% CI)	Optimal cutoff	Sensitivity, % (95% CI)	Specificity, % (95% CI)
GNRI	0.701 (0.636–0.766)	≤ 105.0	51.6 (42.7–60.7)	87.8 (83.1–91.4)
SARC-F	0.651 (0.586–0.712)	≥ 4a	55.7 (46.4–64.4)	64.9 (59.0–70.8)
FRAIL	0.669 (0.604–0.729)	≥ 2b	72.1 (63.6–79.7)	50.6 (44.3–57.0)
MNA-SF	0.675 (0.611–0.734)	≤ 9.5c	42.6 (34.1–51.9)	80.8 (75.6–85.4)

^a Youden-optimal threshold 3.5; expressed as ≥ 4 to reflect integer scoring. ^b Youden-optimal threshold 1.5; expressed as ≥ 2 to reflect integer scoring. ^c On this integer scale, a score of ≤ 9 is categorized as positive for risk. All CIs from BCa bootstrap (5,000 iterations). Extended metrics in Additional file 1: Table S4

Abbreviations: AUC Area under the receiver operating characteristic curve, BCa Bias-corrected and accelerated, CI Confidence interval, FRAIL Frailty screening scale, GNRI Geriatric Nutritional Risk Index, MNA-SF Mini Nutritional Assessment–Short Form, SARC-F Sarcopenia screening questionnaire

Fig. 2.

Receiver operating characteristic curves comparing four geriatric screening tools for predicting 3-year all-cause mortality in the landmark sample (n = 367; 104 events). AUC values with DeLong 95% confidence intervals are displayed in the legend. Open circles indicate Youden-optimal cutoff points. Pairwise DeLong tests versus GNRI: SARC-F, p = 0.213; FRAIL, p = 0.418; MNA-SF, p = 0.529 (all Holm-adjusted p = 1.000; see Table 5)

Table 5.

Pairwise comparisons of area under the receiver operating characteristic curve at 3 years

Comparison	ΔAUC	Z	p-value	p (Holm)
GNRI vs. SARC-F	0.050	1.25	0.213	1.000
GNRI vs. FRAIL	0.032	0.81	0.418	1.000
GNRI vs. MNA-SF	0.026	0.63	0.529	1.000
SARC-F vs. FRAIL	−0.018	−0.65	0.516	1.000
SARC-F vs. MNA-SF	−0.024	−0.60	0.549	1.000
FRAIL vs. MNA-SF	−0.006	−0.19	0.850	1.000

Unadjusted p-values derived using the DeLong test for correlated ROC curves. Holm-adjusted p-values account for multiple testing across six pairwise comparisons

Abbreviations: ΔAUC Difference in area under the curve, FRAIL Frailty screening scale, GNRI Geriatric Nutritional Risk Index, MNA-SF Mini Nutritional Assessment–Short Form, SARC-F Sarcopenia screening questionnaire

Table 6.

Incremental discriminative value of screening tools added to base clinical model

Model	C-statistic	ΔC	p-value
Base (age, sex, eGFR)	0.700	—	—
Base + GNRI	0.744	0.044	< 0.001
Base + SARC-F	0.743	0.043	< 0.001
Base + FRAIL	0.735	0.035	< 0.001
Base + MNA-SF	0.729	0.029	< 0.001

p-values were derived from likelihood ratio tests

Abbreviations: ΔC Change in C-statistic compared to base model, eGFR Estimated glomerular filtration rate, FRAIL Frailty screening scale, GNRI Geriatric Nutritional Risk Index, MNA-SF Mini Nutritional Assessment–Short Form, SARC-F Sarcopenia screening questionnaire

Validation and subgroup robustness

Internal validation via bootstrap resampling confirmed the robust predictive stability of the primary model. Interaction analyses revealed no significant effect modification for GNRI by sex, age, or renal function (Fig. 3).

Fig. 3.

Forest plot of adjusted hazard ratios for all-cause mortality. A Primary multivariable Cox regression model (n = 1,288; 134 events). B Subgroup analyses for GNRI (per unit), stratified by sex, age (< 75 vs. ≥ 75 years), and CKD status. P-values for interaction derived from likelihood ratio tests. The dashed vertical line indicates HR = 1.0

Sensitivity analyses demonstrated that the prognostic significance of GNRI remained consistent after excluding early deaths (< 30 days) and across key clinical subgroups, including patients with and without CKD, both sexes, and age groups above and below 75 years (Additional file 1: Table S5).

Model calibration and reclassification

The base clinical model was well-calibrated. While adding GNRI slightly altered calibration at the extremes of predicted risk, it ultimately reduced the overall prediction error (lowering the Brier score by 9.8% compared to the base model). Importantly, incorporating GNRI into the base clinical model significantly improved the accurate reclassification of patients’ mortality risk (Net Reclassification Improvement: 0.519, p < 0.001; Additional file 2). When applying clinically defined risk thresholds (10% and 30%), this improvement was driven primarily by accurately down-classifying the predicted risk for patients who ultimately survived, highlighting GNRI’s utility in ruling out excess risk in stable patients.

Clinical utility

Decision curve analysis demonstrated that utilizing the GNRI-enhanced model provided a positive net clinical benefit compared with blanket “treat-all” or “treat-none” strategies (Fig. 4). This benefit was most pronounced at intermediate threshold probabilities (20–40%), which is typically where clinical uncertainty regarding nutritional or geriatric interventions is highest.

Fig. 4.

Decision curve analysis comparing the base clinical model (age, sex, eGFR) with the GNRI-enhanced model for 3-year all-cause mortality in the landmark sample (n = 367; 104 events). The 3-year mortality prevalence in the landmark sample was 33.2%

Discussion

In this large outpatient geriatric cohort, the GNRI independently predicted all-cause mortality after adjustment for age, sex, renal function, and concurrent geriatric screening measures, with consistent associations across clinically relevant subgroups. Among the four screening tools evaluated, GNRI achieved the highest numerical standalone discrimination (AUC 0.701), the largest incremental value, and the most favourable reclassification metrics at the primary 3-year endpoint, though pairwise AUC differences did not reach statistical significance after Holm adjustment. Unlike the comparator tools, GNRI requires no patient-reported responses or specialised training, conferring a practical advantage in populations where recall bias, cognitive impairment, or interviewer variability may compromise questionnaire-based screening. These findings support its application as a complementary tool for mortality risk stratification in geriatric outpatient care.

Three key observations emerge from these results. First, the standard GNRI cutoff of ≤ 98 identifies a small (4.0%) but high-risk subgroup with 41.2% mortality, functioning as a highly specific rule-in marker [7, 8, 11]. However, its low sensitivity (17.3%) means that the majority of patients who ultimately die are not flagged. This limitation motivates the data-driven threshold of ≤ 105, which captures a broader at-risk population with improved sensitivity (58.7%) at the cost of modest specificity reduction. In ambulatory geriatric settings, where early identification of borderline nutritional risk is actionable, the ≤ 105 threshold may be more appropriate as the primary screening cutoff. Notably, concordance–discordance analysis revealed that patients flagged by GNRI ≤ 98 but classified as normal by the MNA-SF had mortality equivalent to those flagged by both tools (41.2% vs. 41.2%), whereas patients identified only by the MNA-SF had substantially lower mortality (12.5%), reinforcing that the GNRI captures a biologically distinct dimension of vulnerability not detected by subjective screening.

In our cohort, an at-risk GNRI (≤ 98) was associated with a more than five-fold increase in mortality on univariate analysis (HR 5.45; 95% CI 3.41–8.70). These magnitudes align with the original hospitalised cohort reported by Bouillanne et al. [7] but substantially exceed the pooled estimate (HR 1.32) from a recent meta-analysis [11], likely reflecting the advanced age of our outpatient population relative to the younger, heterogeneous populations in the pooled data.

Second, the complementary diagnostic profiles of the two GNRI thresholds support a tiered approach to nutritional risk stratification. The established cutoff of ≤ 98 functions as a rule-in marker: its high specificity identifies a small subgroup at concentrated mortality risk warranting prompt nutritional intervention and multidisciplinary review. The data-driven threshold of ≤ 105 extends into the Bouillanne “no risk” category (> 98), capturing a borderline population that conventional cutoffs would miss; positive cases would trigger confirmatory assessment via the MNA-SF or comprehensive geriatric evaluation. The graded risk strata further demonstrated a dose–response relationship with mortality, supporting a tiered clinical response calibrated to nutritional risk severity. These workflows are intended to complement, rather than supersede, clinical judgement and concurrent assessment of frailty and sarcopenia. Both thresholds should be considered provisional and are not ready for adoption as hard decision points. Before clinical implementation, local calibration to account for population-specific albumin distributions and nutritional prevalence is required, along with integration with existing tools such as the MNA-SF to ensure that GNRI-based screening complements rather than displaces established assessment workflows. External validation in independent, multicentre cohorts is a prerequisite.

Third, while GNRI demonstrated the highest numerical standalone discrimination among the four tools, pairwise AUC differences did not reach statistical significance after Holm adjustment. Post-hoc power analysis confirmed that the study was underpowered (8–23%) to detect differences of this magnitude, reinforcing the need for adequately powered external comparisons.

This study has several strengths. The sample is large relative to typical outpatient geriatric cohorts, mortality was ascertained through a national registry, and internal validation confirmed the stability of predictive performance. Interaction analyses revealed no significant effect modification by age, sex, or eGFR. The analytical approach incorporated rigorous sensitivity analyses, enhancing methodological transparency.

Several limitations warrant consideration. First, the retrospective, single-centre design restricts causal inference and limits generalisability. As a study from a tertiary geriatric clinic in Türkiye, findings may be influenced by healthcare system structure, referral patterns, and population-specific characteristics (e.g., dietary habits, ethnic composition, comorbidity profiles). The low prevalence of severe nutritional risk (4.0% with GNRI ≤ 98) may reflect the ambulatory setting and may differ in populations with higher baseline malnutrition. Multicentre studies across diverse healthcare systems are needed to confirm the external validity of the proposed thresholds. Residual confounding remains possible, particularly regarding unmeasured inflammatory biomarkers, socioeconomic determinants, and granular measures of disease severity.

Second, the median follow-up of 581 days (1.6 years) was shorter than the primary 3-year endpoint, resulting in a landmark analysis sample of 367 participants (28.5% of the cohort). The 3-year horizon was selected for its clinical relevance to geriatric prognostication, balancing sufficient event accumulation (104 deaths) against excessive censoring. Nevertheless, the data-driven cutoff of ≤ 105 was derived from this landmark sample and should be considered provisional. Predictive performance was consistent at the 1- and 3-year horizons (AUC 0.789 and 0.701, respectively), with 3 years representing the most clinically relevant time frame for geriatric prognostication. The decline in discrimination beyond this horizon further underscores the need for external validation in cohorts with longer follow-up, and the 5-year estimates should be considered exploratory.

Third, the apparent superiority of GNRI over comparator tools was somewhat sensitive to the analytic framework. Finally, statistically optimised cutoffs are inherently susceptible to overfitting in single-cohort studies.

Conclusions

Among geriatric outpatients, the GNRI independently predicts all-cause mortality after adjustment for age, sex, and renal function. Its primary strength lies in its significant incremental value when added to a base clinical model (ΔC-statistic = 0.044, p < 0.001), rather than standalone superiority over existing screens. A data-driven threshold of ≤ 105 for outpatient screening and the established cutoff of ≤ 98 for rule-in support a tiered approach to nutritional risk stratification implementable from routine laboratory data and body weight alone, without reliance on questionnaires or patient cooperation. These thresholds remain provisional pending external validation in multicentre, prospective cohorts with local calibration, and should be integrated with existing tools such as the MNA-SF rather than applied in isolation. Adequately powered studies are needed to determine whether the observed discriminative advantage is clinically meaningful in diverse populations.

Abbreviations

aHR

Adjusted hazard ratio

AUC

Area under the receiver operating characteristic curve

BCa

Bias-corrected and accelerated bootstrap

BMI

Body mass index

CI

Confidence interval

CKD

Chronic kidney disease

CKD-EPI

Chronic Kidney Disease Epidemiology Collaboration

DCA

Decision curve analysis

DOR

Diagnostic odds ratio

eGFR

Estimated glomerular filtration rate

EPV

Events per variable

FRAIL

Fatigue, Resistance, Ambulation, Illnesses, Loss of weight

GNRI

Geriatric Nutritional Risk Index

HR

Hazard ratio

IDI

Integrated Discrimination Improvement

IPCW

Inverse probability of censoring weighting

IQR

Interquartile range

KDIGO

Kidney Disease: Improving Global Outcomes

LOESS

Locally weighted scatterplot smoothing

LR+

Positive likelihood ratio

LR−

Negative likelihood ratio

LRT

Likelihood ratio test

MNA-SF

Mini Nutritional Assessment–Short Form

NRI

Net Reclassification Improvement

ROC

Receiver operating characteristic

SARC-F

Strength, Assistance with walking, Rise from a chair, Climb stairs, Falls

SMD

Standardized mean difference

Authors’ contributions

ML: Conceptualization, Methodology, Software, Formal Analysis, Investigation, Data Curation, Writing—Original Draft, Visualization. SA: Writing—Review and Editing, Supervision, Project Administration. All authors read and approved the final manuscript.

Funding

This research received no external funding. The authors received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

The datasets generated and analyzed during the current study are not publicly available due to institutional and ethical restrictions regarding patient privacy but are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Erciyes University (protocol code 2025/498; date of approval: 22 October 2025). Patient consent was waived by the Ethics Committee of Erciyes University due to the retrospective nature of the study and the exclusive use of de-identified data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.