This was a multicenter retrospective study using data from 2 cohorts. The study was approved by the Ethics Committee of the Affiliated Jinyang Hospital of Guizhou Medical University (Ethical approval number: 2-JY2025). The internal cohort was derived from the Medical Information Mart for Intensive Care IV (MIMIC-IV v3.1) (15): a publicly accessible, open-source repository curated by the Massachusetts Institute of Technology (MIT) comprising 247,366 individuals and 196,527 adults admitted to Beth Israel Deaconess Medical Center between 2008 and 2019. This database was accessed according to an approved protocol, with an Institutional Review Board (IRB) waiver due to retrospective analysis of de-identified data. The external cohort was obtained from patients with cerebral critical illness who were admitted to the Neurological Intensive Care Unit of Jinyang Hospital of Guizhou Medical University from January 2022–January 2025.

This study focused on critically ill neurological patients admitted for the first time to a general intensive care unit (ICU). Diagnoses were classified using the International Classification of Diseases, 9th and 10th revisions (ICD-9/10), including traumatic brain injury, ischemic stroke, intracerebral hemorrhage, subarachnoid hemorrhage, intracranial infections, and related disorders. Inclusion criteria were: (1) age ≥18 years; (2) survival duration and ICU length of stay ≥24 h; (3) availability of complete anthropometric (weight, height) and serum albumin data. For patients with multiple ICU admissions, only data from their first admission were analyzed. Clinical data were collected within the first 24 h following ICU admission.

Extraction was performed using PostgreSQL (v13.7.2) and Navicat Premium (v16.0) with structured query language (SQL). Extracted variables were categorized into six groups:1. Demographics: Age, sex, weight, height, BMI.2. Comorbidities: Hypertension, acute kidney injury (AKI), chronic kidney disease (CKD), heart failure (HF), diabetes, chronic obstructive pulmonary disease (COPD) 0.3. Vital signs: Respiratory rate (RR), heart rate (HR), non-invasive blood pressure (mean: NIBP-M, systolic: NIBP-S, diastolic: NIBP-D) 0.4. Laboratory parameters: Red blood cell count (RBC), white blood cell count (WBC), hemoglobin (Hb), platelet count (PLT), red cell distribution width (RDW), hematocrit (HCT), lactate (Lac), serum sodium (Na), potassium (K), chloride (Cl), creatinine (Cr), blood urea nitrogen (BUN), glucose, international normalized ratio (INR), prothrombin time (PT), activated partial thromboplastin time (PTT), alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin (TB), partial pressure of oxygen (PO₂), partial pressure of carbon dioxide (PCO₂)0.5. Disease severity scores: Acute Physiology Score III (APS III), Simplified Acute Physiology Score II (SAPS-II), Oxford Acute Severity of Illness Score (OASIS), and Sequential Organ Failure Assessment (SOFA) 0.6. Therapies: Vasopressors (VP), mechanical ventilation (MV), continuous renal replacement therapy (CRRT).

Patients were stratified into four nutritional risk groups based on GNRI scores (10): No nutritional risk (GNRI ≥98), Low risk (92 ≤ GNRI <98), Moderate risk (82 ≤ GNRI <92), and High risk (GNRI <82). GNRI was calculated as: GNRI = [14.89 × serum albumin (g/dL)] + [41.7 × (actual BMI/ideal BMI)], with the ideal BMI defined as 22 kg/m² (16, 17). The primary endpoint was 28-day ICU mortality, and the secondary endpoint was 28-day hospital all-cause mortality.

All patients were randomly allocated to training and validation cohorts in a 7:3 ratio. In the training cohort, we performed LASSO regression using the R package “glmnet” with the optimal lambda (λ) value to identify key variables strongly associated with the primary endpoint (28-day ICU mortality). Subsequently, a multivariable Cox regression model was applied to select prognostically significant factors for constructing the predictive model, which was visualized as a nomogram. Model performance was evaluated using receiver operating characteristic (ROC) curves and the area under the curve (AUC).

Continuous variables are presented as mean ± standard deviation (SD) or median with interquartile range (IQR). Group comparisons were performed using t-tests, ANOVA, Mann–Whitney U tests, or Kruskal-Wallis tests, as appropriate. Categorical variables are expressed as numbers (percentages), and differences between groups were analyzed using Pearson’s chi-square or Fisher’s exact tests. Cox proportional hazard regression models were used to estimate hazard ratios (HRs) with 95% confidence intervals (95% CIs). Three models were constructed: Model 1 (unadjusted); Model 2 (adjusted for age, gender, Race, BMI); Model 3 (further adjusted for comorbidities, heart rate, non-invasive systolic blood pressure, vasopressor use, mechanical ventilation, continuous renal replacement therapy, and laboratory parameters differing between survivors and non-survivors; see Supplementary Tables S1–S3). Trend tests and variance inflation factors (VIFs) were assessed to evaluate multicollinearity (Supplementary materials). Restricted cubic splines (RCS) with four knots (5th, 35th, 65th, and 95th percentiles) were applied to explore potential nonlinear relationships between GNRI and outcomes. Subgroup analyses were conducted for predefined strata: age (>65 vs. ≤65 years), BMI (≤30 vs. >30 kg/m²), sex, hypertension, acute kidney injury (AKI), chronic kidney disease (CKD), and COPD. All analyses were performed using R Statistical Software (v4.2.2). A two-sided p-value <0.05 was considered statistically significant.

A total of 1,224 critically ill neurological patients were included based on the eligibility criteria (Table 1). The cohort had a median age of 67 years (IQR: 55–79), with 763 males (62.3%). Stratified by GNRI: 382 patients (31.2%) in the high-risk group, 497 (40.6%) moderate-risk, 194 (15.9%) low-risk, and 151 (12.3%) no-risk. Baseline characteristics stratified by GNRI are shown in Table 1. Patients with nutritional risk (lower GNRI scores) tended to be older. Compared to the no-risk group, high-risk patients showed elevated levels of heart rate, respiratory rate, RDW, WBC, albumin, potassium, lactate, PCO₂, chloride, PTT, creatinine, and urea nitrogen, but lower hemoglobin, red blood cell count, hematocrit, anion gap, and blood pressure. Comorbidities including hypertension, AKI, CKD, heart failure, and COPD were more prevalent in high-risk groups. Clinical severity scores increased significantly with lower GNRI. The high-risk group exhibited substantially higher mortality than the no-risk group: 28-day ICU mortality (31.2% vs. 12.6%, p < 0.001) and overall ICU mortality (35.2% vs. 13.2%, p < 0.001).

As shown in Figure 1, Kaplan–Meier survival curves confirmed a significant inverse correlation between higher GNRI scores and reduced mortality. Lower GNRI groups exhibited significantly higher 28-day ICU mortality (HR = 0.66, 95% CI: 0.55–0.83) and 28-day hospital all-cause mortality (HR = 0.67, 0.54–0.83) compared to higher GNRI groups. Similarly, patients with nutritional risk had markedly elevated mortality versus the no-risk group: 28-day ICU mortality (HR = 1.38, 1.21–1.58) and 28-day all-cause mortality (HR = 1.34, 1.19–1.51).

As shown in Table 2, the multivariable Cox proportional hazards model adjusted for all potential covariates revealed a significant inverse association between higher GNRI scores and 28-day ICU mortality (HR = 0.97, 95% CI: 0.96–0.99). Compared to the no-risk group, patients with moderate and high nutritional risk exhibited increased mortality (HR = 1.66, 1.01–2.74, and HR = 1.79, 1.08–2.97, respectively). Similarly, for 28-day hospital all-cause mortality (Table 3), higher GNRI scores were inversely associated with reduced risk (HR = 0.97, 0.96–0.99). Consistently, moderate- and high-risk groups showed elevated mortality versus the no-risk group (HR = 2.03, 1.26–3.28 and HR = 2.11, 1.30–3.42, respectively). Trend analysis further confirmed a dose–response relationship between escalating nutritional risk and both 28-day ICU/hospital mortality (p < 0.05). RCS Shows Inverse Dose Relationship Between GNRI and Short-Term Mortality in Cerebral Critical Care Patients (nonlinearity p = 0.478 for ICU mortality; p = 0.771 for hospital mortality; Figures 2A,B).

To assess the association between GNRI and 28-day ICU/hospital mortality across demographic and clinical subgroups, we conducted stratified analyses. In the fully adjusted model, results (Supplementary Figures S1, S2) demonstrated that the GNRI-mortality relationship remained consistent regardless of sex, ethnicity, obesity, hypertension, AKI, or CKD. However, this relationship varied by age, diabetes status, and COPD: GNRI showed no statistically significant association with mortality in younger patients (<65 years), those with diabetes, or COPD. Interaction tests confirmed no significant effect modification between GNRI and these subgroups (all p > 0.05).

In the training cohort, 49 candidate variables were analyzed using LASSO regression (Supplementary Figures S3A,B) and multivariable Cox regression. Five independent predictors—GNRI, age, red cell distribution width (RDW), lactate (LAC), and creatinine (CRE)—were selected to construct a risk prediction model for 28-day ICU mortality in critically ill neurological patients (Figure 3). The nomogram demonstrated superior sensitivity and specificity in predicting 28-day ICU mortality compared to traditional severity scores (SOFA, APS III, SAPS II, OASIS), with AUROC values of 0.72 (training cohort; Figure 4A) and 0.69 (validation cohort; Figure 4B).

In an external cohort of 428 neurocritical care patients (28-day mortality rate = 23.60%). Multivariable Cox proportional hazards regression confirmed that higher GNRI scores were significantly associated with reduced 28-day mortality risk (HR = 0.95; 95% CI: 0.93–0.98) (Table 4). Similarly, patients with high nutritional risk demonstrated significantly increased mortality compared to those without nutritional risk (HR = 3.75; 95% CI: 1.21–11.7) (Table 4). Kaplan–Meier curves further demonstrated significantly lower 28-day ICU mortality in patients with GNRI scores above the cohort median (log-rank p < 0.05; HR = 0.67, 95% CI: 0.45–1.00; Figure 5A). Restricted cubic spline analysis revealed a significant inverse dose–response relationship between GNRI and mortality (p = 0.001; Figure 5B). The prognostic nomogram demonstrated robust external validation performance (Figure 5C), with AUC of 0.74 and good calibration between predicted and observed mortality probabilities.