This retrospective cohort study investigated critically ill stroke patients from two independent sources: the MIMIC-IV database and a Chinese institutional intensive care unit (ICU) cohort. Data from MIMIC-IV (version 3.1) covered ICU admissions between 2008 and 2019, whereas the institutional cohort included patients admitted between October 2023 and September 2025.

In the MIMIC-IV cohort, ICU patients with a primary diagnosis of acute stroke were identified using International Classification of Diseases, 9th and 10th Revision (ICD-9/10) codes. Eligible patients were required to be ≥18 years old at ICU admission, to have the first ICU admission during the index hospitalization, and to have documented WBC and blood glucose values within the first 24 h after ICU admission to allow calculation of LGI. Patients with pregnancy, active malignant disease, or missing WBC or blood glucose measurements within the first 24 h after ICU admission were excluded. After applying these criteria, 5,267 adult critically ill stroke patients were included in the final MIMIC-IV cohort. Access to the MIMIC-IV database is granted to individuals who have completed the Collaborative Institutional Training Initiative (CITI) program. One author (Yu Wang; certification number: 70805438) completed the required course and was granted access, which was used for data extraction in this study. The MIMIC-IV database contains de-identified electronic health records from the intensive care units of Beth Israel Deaconess Medical Center (19), and all personal identifiers have been removed and replaced with randomly generated codes; therefore, additional institutional review board approval and individual informed consent were not required.

For the institutional cohort, we retrospectively screened consecutive patients with acute stroke who were admitted to the neurosurgical intensive care unit (NICU) of the First Affiliated Hospital of Chongqing Medical University during the study period. The definition of stroke in the institutional cohort was based on clinicians’ comprehensive assessment of imaging findings and clinical features. Cases were then identified in the electronic medical record system using diagnostic codes and discharge diagnoses consistent with acute stroke. The same eligibility criteria as in the MIMIC-IV cohort were applied: age ≥18 years at NICU admission; first NICU admission during the study period; and availability of WBC and blood glucose measurements within the first 24 h after NICU admission. Patients with pregnancy, active malignant disease, or missing WBC or blood glucose measurements within the first 24 h after NICU admission were excluded. After screening, 424 critically ill stroke patients formed the institutional cohort. This study was approved by the local ethics committee, and the requirement for written informed consent was waived because of its retrospective design and use of anonymized data.

For the MIMIC-IV cohort, data were extracted using PostgreSQL (version 13.7.2) and Navicat Premium (version 15) by running Structured Query Language (SQL) queries. Clinical variables within the first 24 h after ICU admission were collected and included: demographic data (age, sex, and height); comorbidities (hypertension, diabetes mellitus, myocardial infarction, heart failure, renal disease, liver disease, and sepsis); vital signs (heart rate, respiratory rate, systolic, diastolic and mean blood pressure, and body temperature); laboratory parameters (WBC, hemoglobin, platelet count, blood urea nitrogen, serum creatinine, serum sodium, potassium, chloride, calcium, anion gap, bicarbonate, glucose, international normalized ratio [INR], and activated partial thromboplastin time [aPTT]); and ICU severity and treatments, including the Glasgow Coma Scale (GCS) score (20) and the use of vasopressors, insulin therapy, statins, and mechanical ventilation. For the institutional cohort, baseline clinical data were retrospectively retrieved from the hospital electronic medical record system using the same time window (within 24 h of NICU admission) and, as far as possible, the same variable definitions as in the MIMIC-IV cohort. Demographics, comorbidities, vital signs, laboratory measurements, and treatments were collected in a standardized format to ensure comparability between the two cohorts.

The LGI was calculated using the formula “LGI = WBC (×10⁹/L) × blood glucose (mg/dL)/1000” (8, 9). When multiple WBC or glucose measurements were available within the first 24 h after ICU/NICU admission, LGI was calculated using the first available (i.e., first recorded) WBC and the first available blood glucose within that 24-h window; the two values were not required to be from the same blood draw or recorded at the identical time. In both cohorts, patients were categorized into LGI quartiles (Q1–Q4) at admission, with Q1 serving as the reference group for descriptive analyses and Cox models. To address missing data in the MIMIC-IV cohort, variables with more than 20% missingness (height, missing in 54.2% of patients) were excluded from multivariable modeling, whereas variables with 20% or less missingness were handled using multiple imputation. Details of the imputation procedure (variables included, handling of categorical variables, and key settings) are provided in the Statistical analysis section (2.4).

The primary outcome of this study was all-cause mortality within 28 days after ICU admission. The secondary outcomes included all-cause mortality within 365 days and in-hospital mortality. Mortality information in MIMIC-IV was derived from hospital records and validated date-of-death fields provided in the database documentation. For the institutional cohort, all-cause mortality data were extracted from the hospital electronic medical record system and supplemented by follow-up records. Because long-term (365-day) follow-up data were not available for the institutional cohort, only the primary outcome (28-day mortality) was analyzed in this cohort.

Continuous variables were summarized as medians with interquartile ranges (IQR), while categorical variables were presented as counts and percentages. The Kolmogorov–Smirnov test was applied to assess the normality of continuous variables. Group comparisons were performed using the Kruskal–Wallis test for continuous variables and Pearson’s chi-square test or Fisher’s exact test for categorical variables, as appropriate. Baseline characteristics across LGI quartiles were compared separately in the MIMIC-IV and institutional cohorts.

Kaplan–Meier survival curves were constructed to compare the incidence of 28-day and 365-day mortality across LGI quartiles in the MIMIC-IV cohort, and differences among groups were assessed using log-rank tests. For the institutional cohort, Kaplan–Meier curves for 28-day mortality across LGI quartiles were also generated. Cox proportional hazards regression models were used in both cohorts to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the association between LGI and mortality outcomes. In these models, LGI was analyzed both as a continuous variable and as a categorical variable based on quartiles, with Q1 serving as the reference group. Potential confounders were identified based on univariable analyses (p < 0.05) and clinical relevance. Three Cox proportional hazards models were constructed: Model 1 was unadjusted; Model 2 was adjusted for age and sex; Model 3 was further adjusted for GCS score and mechanical ventilation. The extent of missing data for each study variable in the MIMIC-IV cohort is summarized in Supplementary Table 1. In the institutional cohort, there were no missing values for the variables included in the analyses; therefore, no imputation was performed and all analyses were conducted using complete data in the institutional cohort. As described above, in the MIMIC-IV cohort, variables with more than 20% missingness (e.g., height) were not included in the multivariable models. For variables with 20% or less missingness in the MIMIC-IV cohort, multiple imputation by chained equations (MICE) was performed using the mice package (version 3.15.0) in R (21). The imputation model included all variables listed in Table 1 (demographics, vital signs, laboratory tests, comorbidities, and treatments) as well as the outcome variables (28-day and 365-day mortality status and time-to-event) to make the Missing at Random (MAR) assumption more plausible. Predictive Mean Matching (PMM) was used for continuous variables and logistic regression for binary categorical variables. Five imputed datasets were generated, with the imputation process running for 10 iterations per dataset to achieve convergence. The results from analyses conducted on each imputed dataset were pooled using Rubin’s rules to obtain final parameter estimates and standard errors. As a sensitivity analysis, we additionally performed a complete-case analysis in the MIMIC-IV cohort, restricting Models 1–3 to the same complete-case subset defined by complete data on all covariates included in Model 3; results are reported in Supplementary Table 2. The proportional hazards assumption was assessed using Schoenfeld residual plots and Schoenfeld residual-based tests. For the LGI term in the primary 28-day mortality Cox model (Model 3) with LGI entered as a continuous term, the Schoenfeld residual-based test was not significant in either cohort (MIMIC-IV: p = 0.111; institutional cohort: p = 0.794; Supplementary Figure 2).

In the MIMIC-IV cohort, additional analyses were undertaken to describe the shape of the association between LGI and mortality. Restricted cubic spline (RCS) functions with four knots (at the 5th, 35th, 65th, and 95th percentiles of the LGI distribution) were fitted in Cox models for 28-day and 365-day mortality to visualize the dose–response pattern without imposing a specific functional form. In addition, a piecewise Cox proportional hazards model with a single data-driven change point (K) was fitted to provide a parsimonious summary of the association in lower and higher LGI ranges; the estimated change point was treated as a statistical approximation that aids interpretation rather than as a precise biological threshold. Linear and piecewise models were compared using log-likelihood ratio tests. Nonlinear and threshold-effect analyses were not applied to the institutional cohort because of its smaller sample size.

Trend tests across LGI quartiles were performed by entering the quartile number as a continuous variable in the regression model. Prespecified subgroup analyses were conducted in the MIMIC-IV cohort to evaluate the robustness of the association between LGI and mortality across strata of age (<65 vs. ≥65 years), sex, GCS score (≤8 vs. >8), history of myocardial infarction, diabetes mellitus, vasopressor use, insulin therapy, and mechanical ventilation. Interaction terms between LGI and each subgroup variable were examined using likelihood ratio tests to investigate whether the association differed across subgroups. To further assess robustness within the main exposure range, key subgroup analyses were repeated in sensitivity analyses restricted to patients with LGI values below the estimated change points identified in the piecewise Cox models. Interaction analyses were considered exploratory, and no formal adjustment for multiple comparisons was performed. All analyses were conducted using R software (version 4.2.3) and SPSS 22.0 (IBM Corp., Armonk, NY, USA). Two-sided p values < 0.05 were considered statistically significant.

A total of 5,267 adult critically ill stroke patients from the MIMIC-IV cohort and 424 patients from the institutional cohort at a Chinese neurosurgical ICU were included in the final analysis (Figure 1). In both cohorts, baseline characteristics were compared across LGI quartiles (Q1–Q4), as summarized in Table 1 (MIMIC-IV cohort) and Table 2 (institutional cohort). Details of missingness for variables in the MIMIC-IV cohort are provided in Supplementary Table 1. In the institutional cohort, no missing data were observed for the variables analyzed.

In the MIMIC-IV cohort (Table 1), higher LGI quartiles were associated with a progressively more adverse clinical profile. Patients in Q3–Q4 had higher heart rates and respiratory rates and slightly lower systolic, diastolic, and mean blood pressures compared with those in Q1. Laboratory parameters showed a stepwise increase in WBC, blood glucose, blood urea nitrogen, serum creatinine, anion gap, and platelet count across LGI categories, accompanied by lower bicarbonate levels. The prevalence of major comorbidities, including myocardial infarction, heart failure, renal disease, and sepsis, also increased with rising LGI. Consistently, the use of vasopressors, insulin, and mechanical ventilation was more frequent in the higher LGI groups, and patients in Q4 experienced the highest in-hospital, 28-day, and 365-day mortality rates, as well as the longest hospital length of stay.

In the institutional cohort (Table 2), the overall median LGI at admission was 1.13 (IQR 0.80–1.72). Age and sex distribution were broadly similar across LGI quartiles. However, higher LGI was associated with more pronounced physiological and biochemical derangements. Patients in Q3–Q4 had higher WBC counts, blood glucose levels, and anion gap values and lower bicarbonate concentrations compared with those in Q1, and heart rate as well as systolic and mean blood pressure tended to be higher in the upper LGI quartiles. The prevalence of hypertension and diabetes mellitus tended to be higher in the upper LGI quartiles, with the highest proportions observed in Q4. The use of vasopressors and insulin therapy was also more frequent in patients with higher LGI, particularly in Q4. Consistent with these patterns, 28-day mortality, in-hospital mortality, and hospital length of stay were highest in Q4.

Kaplan–Meier survival curves based on LGI quartiles in the MIMIC-IV cohort are presented in Figure 2. The incidence of the primary outcome differed significantly across LGI groups. Over the first 28 days of follow-up, patients with higher LGI values showed progressively lower survival probabilities, with 28-day mortality increasing stepwise from Q1 to Q4 (log-rank p < 0.001; Figure 2A). A similar pattern was observed for long-term outcomes. During 365 days of follow-up, the Kaplan–Meier curves continued to diverge across LGI quartiles, and patients in the highest LGI group (Q4) experienced the poorest survival, whereas those in the lowest LGI group (Q1) had the most favorable prognosis (log-rank p < 0.001; Figure 2B). In the institutional cohort, Kaplan–Meier curves for 28-day mortality across LGI quartiles (Supplementary Figure 1) showed a similar pattern, with progressively lower survival in higher LGI groups.

The proportional hazards assumption was assessed using Schoenfeld residual plots. For the primary 28-day mortality Cox model (Model 3) with LGI entered as a continuous term, the Schoenfeld residual diagnostics suggested no meaningful violation of the proportional hazards assumption in either cohort (Supplementary Figure 2); the Schoenfeld residual-based test for the LGI term was not significant in the MIMIC-IV cohort (p = 0.111) or the institutional cohort (p = 0.794). In the MIMIC-IV cohort, higher LGI levels were consistently associated with increased risks of both 28-day and 365-day all-cause mortality, as summarized in Table 3 and depicted in Figures 3A,B. Crude 28-day mortality increased from 6.69% in Q1 to 23.58% in Q4, and 365-day mortality from 9.73 to 26.46%, with the highest event rates consistently observed in the Q4 group. When LGI was modeled as a continuous variable, each 1-unit increase in LGI was associated with a higher hazard of death in the fully adjusted Cox model (Model 3) for both 28-day mortality (HR 1.179, 95% CI 1.150–1.209, p < 0.001) and 365-day mortality (HR 1.169, 95% CI 1.142–1.197, p < 0.001). Compared with patients in the lowest LGI quartile (Q1), those in higher quartiles had progressively elevated risks of death. In Model 3, the HRs for 28-day mortality were 1.244 (95% CI 0.945–1.638) for Q2, 1.834 (95% CI 1.420–2.367) for Q3, and 3.259 (95% CI 2.568–4.138) for Q4 (p for trend < 0.001). A similar graded pattern was observed for 365-day mortality, with HRs of 1.143 (95% CI 0.904–1.445) for Q2, 1.561 (95% CI 1.252–1.946) for Q3, and 2.703 (95% CI 2.203–3.317) for Q4 (p for trend < 0.001). In a complete-case sensitivity analysis of the MIMIC-IV cohort (n = 3,764; 71.46% of the original cohort), the association between LGI and mortality remained consistent (Supplementary Table 2). In the fully adjusted model (Model 3), the HRs for 28-day mortality across LGI quartiles were 1.288 (95% CI 0.933–1.777) for Q2, 1.894 (95% CI 1.406–2.551) for Q3, and 3.020 (95% CI 2.277–4.006) for Q4 (p for trend < 0.001). Similarly, the HRs for 365-day mortality were 1.147 (95% CI 0.878–1.500) for Q2, 1.502 (95% CI 1.166–1.935) for Q3, and 2.414 (95% CI 1.904–3.061) for Q4 (p for trend < 0.001).

In the institutional cohort, the direction of association between LGI and 28-day mortality was concordant with that observed in the MIMIC-IV cohort (Table 3). In the fully adjusted model (Model 3), LGI as a continuous variable remained significantly associated with 28-day mortality (HR 1.351, 95% CI 1.090–1.675, p < 0.05). Using Q1 as the reference, patients in Q4 had a markedly higher risk of 28-day death (HR 4.374, 95% CI 1.270–15.063), while HRs for Q2 and Q3 were 1.542 (95% CI 0.367–6.477) and 1.646 (95% CI 0.419–6.466), respectively (p for trend = 0.004). Although the CIs were wider than in the MIMIC-IV cohort—reflecting the smaller sample size and fewer events—the overall pattern supported a positive, dose–response relationship between LGI and short-term mortality in the external cohort. Because long-term (365-day) follow-up data were not available for the institutional cohort, only the primary outcome of 28-day all-cause mortality was evaluated in this cohort. In the institutional cohort, quartile-based estimates warrant cautious interpretation given the relatively small sample size and limited events.

To further describe the association between LGI and mortality, RCS analyses were performed in the MIMIC-IV cohort (Figures 3C,D). These analyses revealed a positive but nonlinear association between LGI and mortality, with risk rising steeply from low to moderate LGI levels and then increasing more gradually at higher levels. In the piecewise Cox proportional hazards models with a single change point, the estimated inflection points were 2.89 for 28-day mortality and 3.00 for 365-day mortality (Table 4). Below these points, LGI showed a strong association with mortality (28-day mortality: HR 1.95, 95% CI 1.77–2.15; 365-day mortality: HR 1.72, 95% CI 1.58–1.88; both p < 0.001), whereas above these inflection points the slope of the association was markedly attenuated (28-day mortality: HR 1.05, 95% CI 1.00–1.11, p = 0.045; 365-day mortality: HR 1.06, 95% CI 1.02–1.11, p = 0.009). The log-likelihood ratio tests indicated that the two-segment models provided a significantly better fit than a single linear term for LGI (both p < 0.001).

To evaluate the robustness of the association between LGI and mortality and to explore potential effect modification, subgroup analyses with interaction tests were performed in the MIMIC-IV cohort, stratified by age, sex, GCS score, history of myocardial infarction, diabetes mellitus, vasopressor use, insulin therapy, and mechanical ventilation status (Figures 4, 5). For 28-day mortality, higher LGI was consistently associated with an increased risk of death across all strata, with HR remaining above 1.0 in each subgroup. The strength of this association varied, with statistically significant effect modification observed for age (p for interaction < 0.001), sex (p for interaction = 0.009), diabetes mellitus (p for interaction = 0.006), and insulin therapy (p for interaction = 0.048). In general, the association between LGI and 28-day mortality appeared stronger among younger patients (<65 years), men, individuals without diabetes, and those receiving insulin therapy. No significant effect modification was detected for GCS score, myocardial infarction, vasopressor use, or mechanical ventilation status with respect to 28-day mortality (all p for interaction > 0.05).

For 365-day all-cause mortality, elevated LGI was likewise associated with higher risk in all examined subgroups. Significant effect modification was again noted for age (p for interaction = 0.001), sex (p for interaction = 0.006), and diabetes mellitus (p for interaction = 0.029), with a generally stronger association in younger patients, men, and those without diabetes, whereas no statistically significant effect modification was found for GCS score, myocardial infarction, vasopressor use, mechanical ventilation, or insulin therapy (all p for interaction > 0.05).

Because the number of patients with very high LGI values above the estimated change points from the threshold-effect analysis (approximately 2.89 for 28-day mortality and 3.00 for 365-day mortality) was limited, additional sensitivity analyses were conducted by restricting the MIMIC-IV cohort to patients with LGI values below these change points (Supplementary Figures 3, 4). Within this main exposure range, the overall positive association between LGI and both 28-day and 365-day mortality remained evident, and in many subgroups the HRs continued to suggest higher risk with increasing LGI, although in some strata the estimates were closer to unity or even slightly below 1.0. For 28-day mortality, the effect modification by age, sex, diabetes mellitus, and insulin therapy that was statistically significant in the primary analyses was no longer statistically significant in the restricted analyses, whereas interactions with GCS score and vasopressor use reached nominal significance. For 365-day mortality, after restriction to LGI values below the change point, interaction terms for age and sex were no longer statistically significant, whereas the interaction with diabetes mellitus remained significant, with the direction of effect differing from the primary analysis, and interaction p values for insulin therapy and vasopressor use were also < 0.05. These shifts in p for interaction likely reflect the reduced number of events and narrower exposure range in the restricted cohort and therefore should be interpreted cautiously. Overall, despite some inconsistencies in interaction patterns, the subgroup and sensitivity analyses support LGI as a robust prognostic marker in critically ill stroke patients, while suggesting potential heterogeneity of effects across demographic and metabolic strata. All interaction analyses were considered exploratory, and no formal adjustment for multiple comparisons was performed.