The data for this retrospective study were derived from the clinical follow-up data of centenarians and older adults recorded in Hainan Hospital, Chinese PLA General Hospital, from January 2015 to December 2023, combined with the household registration system and death registries of community health service centers. Data collection was conducted in strict accordance with the ethical principles of the Declaration of Helsinki and was approved by the Ethics Committee of Hainan Hospital, Chinese PLA General Hospital (Approval No. 301HN11201601). Written informed consent was obtained from all participants or their legally authorized representatives prior to data collection. The death outcomes during follow-up were verified through the household registration system and community health service centers. Data management adopted double-entry and cross-checking procedures to ensure completeness and accuracy, in accordance with the Standard Protocol Items: Recommendations for Interventional Trials (SPIRIT). The study posed minimal risk and ensured participant privacy and confidentiality through encrypted data storage compliant with the 2023 version of the Health Insurance Portability and Accountability Act (HIPAA).

The inclusion criteria were centenarians (aged ≥100 years) admitted to Hainan Hospital, Chinese PLA General Hospital who had GNRI measured during the index hospitalization and complete mortality follow-up. Hospitalization was used to define the baseline assessment time point, and mortality outcomes were obtained during subsequent follow-up via official registries. To maintain the integrity of the study and ensure the robustness of its results, strict exclusion criteria were implemented: (1) individuals aged less than 100 years were excluded; (2) patients with missing GNRI indicators; (3) individuals diagnosed with advanced-stage renal impairment, hepatic cirrhosis, or malignancies; (4) patients who did not receive follow-up after discharge; Finally, a total of 1,002 centenarians were included and divided into four groups according to the quartiles of GNRI (Figure 1). We selected quartiles a priori to provide a balance between granularity and statistical stability, yielding comparable group sizes and sufficient events per group for robust Cox regression estimates. Quartile-based categorization is also commonly used in nutritional epidemiology and facilitates comparison with prior studies.

Data extraction is carried out by professional clinicians. The extraction of potential variables could be divided into four main groups: (1) demographics, such as age, gender, nationality, marital status, education, BMI. (2) Comorbidities, including diabetes mellitus, hypertension. (3) Diet and Lifestyle, including diet rules, fullness level, number of meals per day, the types of food consumed in the past week, smoking, drinking alcohol. (4) Nutrition and Outcomes, including GNRI, GNRI group, life state, time.

The variable ‘Time’ is defined as the follow-up duration, calculated as the interval from the date of hospital admission to the date of death or the date of the last follow-up interview. To ensure precision in the survival analysis, this duration was recorded in months. The calculation formula for GNRI: GNRI = 1.489 × albumin (g/L) + 41.7 × (body weight/ideal body weight) (17), the ideal body weight is defined as [height (m)]² × 22. All dietary and lifestyle variables and nutritional and outcome variables were extracted from the complete follow-up data of the patients. In this study, Ideal Body Weight (IBW) was calculated based on the standard Body Mass Index (BMI) target of 22 kg/m², using the formula: IBW = 22 × [height (m)]². This method was chosen for its applicability and widespread use in nutritional assessment for the elderly population; in this study, IBW was used as a standardized reference weight (rather than an age-specific target weight for centenarians), and the present-to-ideal weight ratio was truncated at 1 when present weight exceeded IBW, as recommended in the original GNRI method (11).

To avoid possible bias, variables were excluded if they had more than 20% missing values. Variables with missing data less than 20% were processed by multiple imputation using a random forest algorithm (trained by other non-missing variables) by the “mice” package of R software (18, 19).

The endpoint of this study is all—cause mortality.

Continuous variables are reported as mean ± standard deviation (SD) or median and interquartile range (IQR), contingent upon the data distribution while representing the categorical variables as frequencies and percentages. Group comparisons for continuous variables with normal distribution were performed using Welch’s t-test or ANOVA, and non-normally distributed variables were compared using the Wilcoxon rank-sum test or Kruskal–Wallis test. For comparison between groups of categorical data, we used the Fisher exact test for expected frequencies <5; otherwise, we used the Chi-squared test. To further explore specific differences between groups, Tukey’s Honest Significant Difference (HSD) post-hoc test was employed following significant results from one-way ANOVA. Additionally, considering the ordinal nature of the GNRI quartiles, the Cochran-Armitage trend test was utilized to assess the presence of monotonic trends in categorical variables across the groups Cox proportional hazards models were used to calculate the hazard ratio (HR) and 95% confidence interval (CI) between the GNRI and endpoints, and also adjusted for some models. The proportional hazards assumption was verified using the scaled Schoenfeld residuals test. A global test p-value greater than 0.05 was considered to indicate no violation of the assumption. Confounding variables included variables selected based on p value <0.05 in univariate analysis. And clinically relevant and prognosis-associated variables were also enrolled in the multivariate model: model 1: unadjusted; model 2: adjusted for age, BMI, gender, nationality, marital status, education, smoking, and drinking alcohol; model 3: adjusted for age, BMI, number of meals per day, gender, nationality, marital status, education, drinking alcohol, smoking, diabetes mellitus, hypertension, diet rules, fullness level, red meat, poultry, fish and seafood, eggs, dairy products, legumes, nuts, vegetable, and fruit. To ensure the reliability of the multivariable regression analysis (and avoid result bias caused by multicollinearity among covariates), multicollinearity diagnosis was performed for all included variables. Further, we also analyzed the nonlinear association between baseline GNRI and all-cause mortality using a restricted cubic spline regression model. The threshold effect of GNRI was identified using the piecewise Cox regression mode. Further stratified analyses were performed based on gender, nationality, marital status, education, smoking, drinking alcohol, diabetes mellitus, hypertension, diet rules, fullness level and the types of food consumed in the past week to identify the consistency of the prognostic value of the GNRI for primary outcomes. The interactions between GNRI and variables used for stratification were examined with likelihood ratio tests.

Kaplan–Meier survival analysis was employed to assess the incidence rate of endpoints among groups based on different levels of the GNRI, and their differences were assessed through log-rank tests.

To enhance the reliability of the results, 1:1 propensity score matching was performed for the included participants. Specifically, we employed a nearest-neighbor matching algorithm without replacement. A caliper width of 0.05 was applied to strictly match participants. This parameter was selected to ensure that the standardized mean difference (SMD) for all covariates was reduced to less than 0.1, thereby maximizing the balance between the groups. Based on the post-matching results, we further conducted the following analyses: (1) Cox proportional hazards models were employed to calculate the hazard ratio (HR) and 95% confidence interval (CI) for the association between GNRI and the endpoint; (2) restricted cubic spline regression models were used to analyze the nonlinear relationship between baseline GNRI and all-cause mortality; (3) piecewise Cox regression models were applied to identify the optimal threshold effect of GNRI; (4) additional subgroup analyses were performed to determine the consistency of the prognostic value of GNRI for the primary outcome; and (5) Kaplan–Meier survival analysis was utilized to assess the incidence of the endpoint across different groups. A double-sided p < 0.05 was regarded as statistically significant. All statistical analysis was performed by the R software (version 4.2.2).

In this study, a total of 1,002 centenarians were enrolled, with a median age of 102 years (IQR: 101–104). The cohort included 180 men (18.0%) and 822 women (82.0%). Over a median follow-up of 3 years, the cumulative all-cause mortality was 92.7%. The baseline characteristics of the study population stratified by GNRI quartiles are presented in Table 1. Participants in higher GNRI quartiles were characterized by significantly higher BMI, longer follow-up duration, and a higher proportion of males (all p < 0.05). Post-hoc comparisons confirmed that the highest GNRI quartile (Q4) had significantly greater BMI than all other groups and longer follow-up time compared to the lowest quartiles. Notably, survival rates demonstrated a significant monotonic increase across GNRI quartiles (Cochran–Armitage trend test, p < 0.001), rising from 1.8% in the lowest quartile (Q1) to a peak of 10.2% in Q3, before plateauing in Q4 (9.9%). Significant differences were also observed in education levels, smoking status, and hypertension prevalence across the groups. Detailed dietary consumption patterns across the GNRI quartiles are now presented in Supplementary Table S1. In summary, participants in higher GNRI quartiles demonstrated a significantly higher frequency of consuming protein-rich foods, including red meat, poultry, fish, eggs, and dairy products (all p < 0.05), indicating a positive correlation between nutritional status and dietary diversity. Table 2 summarizes the comparison of demographic and clinical factors between survivors and non-survivors. Non-survivors were significantly more likely to be female, younger, and had lower BMI compared to survivors. The baseline distribution of specific dietary components categorized by life state is presented in Supplementary Table S2, which generally indicated better dietary intake among survivors. Crucially, the GNRI levels were significantly lower in the non-survivor group compared to the survivor group (92 ± 11 vs. 96 ± 8, p < 0.001).

Cox proportional hazards analysis demonstrated that GNRI was a significant independent predictor of all-cause mortality across all models (Table 3). When analyzed as a continuous variable in the fully adjusted model 3, each 1-unit increase in GNRI was associated with a 3% reduction in the hazard of mortality [HR 0.97 (95% CI 0.96–0.98), p < 0.001]. Similarly, when categorized into quartiles, a clear graded protective effect was observed. Participants in the highest quartile (Q4) exhibited a 52% lower risk of death compared to the lowest quartile [HR 0.52 (95% CI 0.41–0.66)], with a highly significant linear trend across the groups (p for trend <0.001). The assessment of the proportional hazards assumption based on Schoenfeld residuals yielded a global p-value of 0.12, indicating that the assumption was generally met. Although visual inspection of the survival curves suggested potential crossing or convergence in the late follow-up period, statistical tests did not confirm a significant global violation. Additionally, collinearity diagnostics indicated no significant multicollinearity among the included covariates (all VIF < 1.8, Supplementary Table S3), ensuring the stability of the regression coefficients.

Restricted cubic spline analysis revealed a significant non-linear relationship between GNRI and mortality (p for non-linearity <0.001, Figure 2), characterized by a steep decline in risk at lower GNRI values followed by a saturation effect. Piecewise Cox regression identified an optimal threshold at GNRI = 96.378 (Table 4). Below this threshold, GNRI was significantly inversely associated with mortality, with a 4% risk reduction per unit increase [HR 0.96 (95% CI 0.95–0.97), p < 0.001]. However, above this threshold, the protective benefit plateaued, and no further significant reduction in mortality was observed [HR 1.01 (95% CI 0.99–1.02), p = 0.443].

Figure 3 presents the subgroup analyses stratified by demographic, clinical, and dietary factors. The inverse association between GNRI and mortality remained robust across most strata, including nationality, marital status, and various dietary habits (detailed in Figure 3), indicating that the prognostic value of GNRI is independent of these covariates. A significant interaction was observed only for gender (p for interaction = 0.001). The protective effect of GNRI was highly prominent in females [HR 0.96 (95% CI 0.95–0.97)] but was not statistically significant in males [HR 1.01 (95% CI 0.99–1.03)]. Kaplan–Meier survival curves (Figure 4) visually confirmed these findings, showing distinct separation between the groups, where centenarians in lower GNRI quartiles experienced significantly higher cumulative mortality rates compared to those in higher quartiles.

To ensure the robustness of our findings and minimize selection bias, propensity score matching (PSM) was performed to balance baseline covariates between groups. The distribution of propensity scores was highly similar between the matched groups. As shown in Supplementary Table S4, Supplementary Figure S1, standardized mean differences (SMDs) for all matched covariates were significantly reduced to below 0.1, confirming that the matching process effectively eliminated baseline imbalances. Sensitivity analyses using the matched cohort yielded results consistent with the primary analysis (Table 5). GNRI remained a significant independent predictor of mortality, both as a continuous variable (HR 0.98 per unit increase, p < 0.001) and when categorized into quartiles, showing a graded protective effect (p for trend <0.001). Specifically, participants in the highest GNRI quartile (Q4) exhibited a 42% lower risk of mortality compared to the lowest quartile [HR 0.58 (95% CI 0.43–0.80)]. Restricted cubic spline analysis further confirmed the significant non-linear relationship (p for non-linearity <0.001, Figure 5), characterized by a steep reduction in mortality risk as GNRI increased up to approximately 96, followed by a plateau phase.

Piecewise Cox regression validated the optimal threshold at GNRI = 96.378 (Table 6, Supplementary Figure S2). Consistent with the pre-matching analysis, a significant inverse association was observed below this threshold (HR 0.95, p < 0.001), indicating a 5% risk reduction per unit increase in this range. Conversely, no significant further benefit was seen above the threshold (HR 1.02, p = 0.136), suggesting a saturation of the survival benefit.

Subgroup analyses in the matched cohort (Supplementary Figure S3) aligned with the primary findings, identifying a significant interaction with gender (p for interaction = 0.002). The protective effect of GNRI was statistically significant in females [HR 0.97 (95% CI 0.96–0.98)] but was not observed in males [HR 1.01 (95% CI 0.99–1.03)]. No significant interactions were found for other sociodemographic or lifestyle variables. Finally, Kaplan–Meier curves (Supplementary Figure S4) confirmed that higher GNRI levels were associated with significantly better long-term survival. Overall, the consistency of results across Cox regression, threshold analysis, and subgroup stratification after propensity score matching strongly reinforces the validity of our primary conclusions.