Association between glycemic variability and all-cause mortality in critically ill patients with non-traumatic subarachnoid hemorrhage: a retrospective study based on the MIMIC-IV database

Abstract

Background

Abnormal glycemic variability (GV), defined as acute fluctuations in blood glucose, is a prevalent phenomenon observed in critically ill patients and has been linked to unfavorable outcomes, including elevated mortality. However, the impact of this factor on patients with non-traumatic subarachnoid hemorrhage (SAH) remains unclear. The aim of this study is to explore the relationship between GV and all-cause mortality (ACM) in patients with non-traumatic SAH.

Methods

All blood glucose measurements taken within the initial 72-h period following intensive care unit (ICU) admission for non-traumatic SAH patients were extracted. The coefficient of variation (CV) was employed to quantify GV, defined as the ratio of the standard deviation (SD) to the mean blood glucose. Patients were stratified into tertiles based on their GV. Furthermore, we assessed ACM at multiple timepoints, including at ICU, in-hospital, 30 days, 90 days, 180 days, and 1 year. The relationship between GV and ACM was analyzed using Cox proportional hazards regression models and restricted cubic splines (RCS). Kaplan–Meier survival curves were used to estimate survival across different GV groups. Subgroup analyses were performed to evaluate the robustness of the findings.

Results

The study cohort comprised a total of 1056 patients, of whom 55.6% were female. The mortality rates observed in the ICU, hospital, and at various timepoints, including 30 days, 90 days, 180 days, and 1 year, were 12.8%, 16.2%, 17.5%, 21.5%, 24.3%, and 26.6%, respectively. Multivariate Cox regression analysis revealed a significant association between the high GV (≥ 20.4%) and ACM among patients with SAH. RCS analysis revealed a nonlinear U-shaped correlation between GV and ACM.

Conclusions

GV was identified as an independent risk factor for ACM in critically ill patients with non-traumatic SAH. These findings indicate that enhancing GV stability could potentially contribute to reducing mortality rates among non-traumatic SAH patients.

Introduction

Spontaneous subarachnoid hemorrhage (SAH) constitutes approximately 5% of all stroke cases and represents the third most prevalent subtype of cerebrovascular disease. SAH is strongly associated with long-term cognitive deficits, significantly impaired quality of life, and increased mortality risk [1]. According to epidemiological evidence, SAH is associated with a case fatality rate of 30% to 50%. In addition, more than one-fifth of patients who survive the initial event do not recover functional independence, underscoring the significant disability burden imposed by this condition [2, 3]. With the global aging population trend, stroke-related demands on intensive care unit (ICU) are escalating markedly. This escalation underscores the imperative need to identify and validate prognostic biomarkers that predict unfavorable outcomes in stroke patients. Such biomarkers should ideally feature simplicity, clinical utility, cost-efficiency, and seamless integration into routine care protocols.

Recent studies have shown that stringent glycemic control does not improve clinical outcomes or reduce mortality in patients with acute stroke. Instead, it may increase the incidence of hypoglycemia, as demonstrated in a recent randomized trial. This counterintuitive finding highlights the need for personalized glucose management strategies in neurocritical care [4, 5]. Similar observations have been reported in other patient populations [6, 7]. Glycemic variability (GV), defined as the oscillations in blood glucose levels beyond the normal range over a specific period, is a pivotal factor in evaluating the efficacy of glycemic control. Elevated GV constitutes a primary manifestation of dysglycemia in critically ill patients and serves as a prognostic marker for adverse outcomes across diverse disease entities [8,9,10]. In vitro experiments have demonstrated that, compared to stable hyperglycemia, significant GV is associated with increased endothelial dysfunction and the induction of oxidative stress, potentially leading to more severe cerebrovascular damage [11]. However, the role of GV in the clinical management of cerebrovascular diseases has garnered limited attention, particularly regarding its application in SAH. Consequently, this study, utilizing data from the Medical Information Mart for Intensive Care (MIMIC-IV) database, investigates the association between GV and all-cause mortality (ACM) in critically ill patients with non-traumatic SAH. This research may aid physicians in identifying high-risk patients, thereby facilitating more informed medical decision-making and enabling more precise monitoring or timely treatment of these vulnerable individuals.

Materials and methods

Data source

Following completion of the National Institutes of Health training requirements for human subjects protection and successful passage of Collaborative Institutional Training Initiative Program certification (certification number: 50289289), authorized researcher Yuyang Hou extracted study data from the MIMIC-IV database. This retrospective cohort study utilized the publicly available critical care database containing deidentified clinical records of ICU and emergency department patients at Beth Israel Deaconess Medical Center (Boston, MA, USA) from 2008 to 2022. The database comprehensively documents demographic characteristics, laboratory parameters, pharmacological interventions, physiological measurements, surgical procedures, diagnostic classifications, and longitudinal survival outcomes. The research protocol was conducted in accordance with the Declaration of Helsinki’s ethical principles, with individual informed consent waived due to the pre-existing anonymized nature of MIMIC-IV data.

Study population

Patients with non-traumatic SAH were identified using International Classification of Diseases (ICD)−9 code 430 and ICD-10 codes I60, I600–I6012, I6000–I6002, I6020–I6022, I6030–I6032, and I6050–I6052. For subjects with multiple ICU admissions, only the initial ICU admission was analyzed. Exclusion criteria comprised: (1) age < 18 years (n = 0), (2) ICU length of stay < 6 h (n = 4), (3) fewer than three blood glucose measurements within the first 72 h of ICU admission (n = 238), and (4) critical data deficiencies (n = 1). The final analytic cohort comprised 1056 patients meeting inclusion criteria. The patient inclusion flowchart is displayed in Fig. 1.

Fig. 1

Flowchart of participant selection. SAH, subarachnoid hemorrhage; ICU, intensive care unit; GV, glycemic variability

Data extraction and definitions

The baseline reference timepoint was defined as the first calendar day of ICU admission. Clinical variables encompassing demographics, comorbidities, severity score, vital signs, laboratory profiles, and therapeutic interventions were extracted through structured queries developed using the MIMIC-IV code repository (Massachusetts Institute of Technology Laboratory for Computational Physiology; https://github.com/MIT-LCP/mimic-code). Data extraction was performed via Navicat Premium (v17.0.8) using SQL scripting [10, 12, 13]. Missing data handling followed a two-stage protocol: variables exceeding 20% missingness were excluded, while remaining incomplete variables (≤ 20% missing) underwent multiple imputation via chained equations with 10 imputed data sets using predictive mean matching (5 iterations, random seed = 123). The first imputed data set was retained for primary analysis after validation of convergence diagnostics. In the absence of a gold standard for GV assessment, we utilized the coefficient of variation (CV) as the primary metric, a methodology widely adopted in critical care research [14,15,16]. The CV was calculated using the following formula: standard deviation (SD) divided by the mean, expressed as a percentage (SD/mean × 100). This approach stands in contrast to metrics, such as the Mean Amplitude of Glycemic Excursions (MAGE), which requires continuous glucose monitoring (CGM) data at 5-min intervals—a resolution unattainable in our ICU cohort. The selection of CV over MAGE was further justified by its proven feasibility in neurosurgical ICU, where intracranial pressure monitoring devices often preclude CGM electrode placement, and by the prohibitive costs of continuous monitoring in large-scale retrospective studies. By restricting the analysis to patients with at least three blood glucose measurements, we aimed to enhance the validity and generalizability of our findings. This approach ensures that the GV assessment is based on a sufficient number of data points, thereby providing a meaningful and stable estimate.

Clinical outcomes

The endpoints of the current study were ICU, in-hospital, 30 days, 90 days, 180 days and 1-year ACM. Details regarding deaths were obtained from state and hospital records. The follow-up for all individuals in the MIMIC-IV database extended to at least 1 year.

Statistical analysis

For continuous variables, the mean ± SD or median and interquartile range (IQR) were applied as appropriate to represent the data, in accordance with the characteristics of their distribution. Between-group comparisons employed Student's t tests or ANOVA for normally distributed data and Mann–Whitney or Kruskal–Wallis tests for skewed distributions. Categorical variables were expressed as frequencies with percentages and analyzed using chi-square or Fisher’s exact tests. Given the non-normal distribution of the CV, it was analyzed as a continuous variable and further stratified into tertiles (T1–T3) for risk stratification, a method widely adopted in GV research.

Cox proportional hazards regression analysis was conducted to evaluate the associations between GV and mortality. To account for potential confounding factors, the analysis was stratified across several distinct models. The crude model did not include any adjustments. Model 1 was adjusted for demographic characteristics, including age, sex, and ethnicity. Building on Model 1, Model 2 further adjusted for physiological parameters and treatments, such as heart rate, mean arterial blood pressure, oxygen saturation, continuous renal replacement therapy (CRRT), mechanical ventilation (MV), antidiabetic therapy, and Glasgow Coma Scale (GCS). Model 3 expanded on Model 2 by incorporating additional adjustments for comorbidities, including hypertension, diabetes, myocardial infarction, congestive heart failure, chronic obstructive pulmonary disease, peripheral vascular disease, and renal diseases.

Kaplan–Meier survival analysis with log-rank testing was employed to visualize and compare ACM risk across GV tertiles (T1–T3). To assess potential nonlinear associations, we implemented restricted cubic splines (RCS) with four knots positioned at the 5th, 35th, 65th, and 95th percentiles of the GV distribution. Interaction effects were evaluated through stratified subgroup analyses, with results graphically represented via forest plots displaying hazard ratios (HR) and 95% confidence intervals (CI). All statistical analyses were performed using R software version 4.3.2 (http://www.r-project.org). P value < 0.05 was considered statistically significant.

Results

Patient characteristics

This study included 1056 critically ill patients with non-traumatic SAH, with a median age of 61 years (IQR: 50–72). The cohort comprised 54.4% females (n = 587) and was predominantly White (57.6%, n = 608). Mortality rates demonstrated a temporal escalation: 12.8% during ICU admission, 16.2% during hospitalization, and 17.5%, 21.5%, 24.3%, and 26.6% at 30 days, 90 days, 180 days, and 1-year post-admission, respectively. Participants were stratified into GV tertiles based on CV thresholds: T1: CV < 14.2%; T2: 14.2 ≤ CV < 20.4%; T3: CV ≥ 20.4%.

Table 1 details baseline demographics and clinical characteristics across GV tertiles. Patients in the highest CV tertile (T3: CV ≥ 20.4%) demonstrated a higher prevalence of metabolic and cardiovascular comorbidities, including diabetes mellitus, myocardial infarction, congestive heart failure, and renal disease, compared to lower GV groups. Physiological and laboratory profiles revealed elevated markers of systemic stress in high-GV patients, including increased heart rate, leukocyte counts, serum creatinine, blood urea nitrogen, prothrombin time (PT), and international normalized ratio (INR). Disease severity scores consistently indicated greater illness burden in this group, with higher Sequential Organ Failure Assessment (SOFA), Simplified Acute Physiology Score II (SAPS II), Acute Physiology Score III (APS III), Oxford Acute Severity of Illness Score (OASIS), and Systemic Inflammatory Response Syndrome (SIRS) scores. Therapeutic intensity was correspondingly elevated in high-GV patients, with increased utilization of MV, CRRT, and glucose-lowering agents. Mortality rates demonstrated a graded association with GV levels, with the highest CV tertile showing significantly elevated ACM across all timepoints: ICU admission (19.6% vs. 7.7% in T1), hospitalization (24.1% vs. 10.8%), and post-discharge intervals (30 days: 24.4% vs. 12.8%; 90 days: 29.8% vs. 15.3%; 180 days: 33.0% vs. 17.6%; 1 year: 35.2% vs. 19.6%).

Table 1 Characteristics and outcomes of participants according to tertiles of glycemic variability

Clinical outcomes

Figure 2 presents the temporal patterns of ACM across CV tertiles using Kaplan–Meier survival curves. Patients in the highest CV tertile (T3) demonstrated significantly reduced survival probabilities at all evaluated timepoints compared to lower GV groups (log-rank P < 0.05). Cox proportional hazards models were implemented to quantify GV-associated mortality risk during ICU admission, hospitalization, and at 30 days, 90 days, 180 days, and 1-year intervals post-admission. The corresponding HR with 95% CI are detailed in Table 2.

Fig. 2

Kaplan–Meier survival analysis curves of ACM by GV A ICU mortality; B in-hospital mortality; C 30-day mortality; D 90-day mortality; E 180-day mortality; F 1-year mortality. ACM all-cause mortality, GV glycemic variability

Table 2 Multivariate Cox regression analysis for morality in non-traumatic SAH patients

When analyzed as a continuous variable, each unit increase in GV was associated with a 4% elevated risk of ICU mortality in the crude model (HR: 1.04; 95% CI 1.03–1.05; P < 0.001). This association remained significant across progressively adjusted models: Model 1 (HR: 1.04; 95% CI 1.02–1.05; P < 0.001), Model 2 (HR: 1.03; 95% CI 1.02–1.04; P < 0.001), and Model 3 (HR: 1.03; 95% CI 1.02–1.05; P < 0.001). When analyzed as a categorical variable, patients in the highest tertile of CV (T3) had a significantly higher risk of ACM in the ICU compared to those in T1, with a HR of 2.12 (95% CI 1.35–3.31; P = 0.001) in the crude model. This significance persisted in Model 1 (HR: 2.00; 95% CI 1.27–3.15; P = 0.003), Model 2 (HR: 1.64; 95% CI 1.02–2.63; P = 0.043), and Model 3 (HR: 1.66; 95% CI 1.01–2.73; P = 0.044). These findings, whether considering CV as a continuous or categorical variable, consistently demonstrate a significant relationship with ACM, suggesting that CV may be considered a robust predictor of ACM within the ICU environment. For in-hospital mortality, a continuous increase in CV was significantly associated with ACM in the crude model (HR: 1.03; 95% CI 1.02–1.04; P < 0.001) and across all adjusted models (Model 1: HR: 1.03; 95% CI 1.02–1.04; P < 0.001; Model 2: HR: 1.03; 95% CI 1.01–1.04; P < 0.001; Model 3: HR: 1.03; 95% CI 1.01–1.04; P < 0.001). However, categorical analysis showed that while T3 exhibited a higher risk in the crude model (HR: 1.69; 95% CI 1.14–2.49; P = 0.008) and Model 1 (HR: 1.65; 95% CI 1.11–2.44; P = 0.013), significance was lost in Model 2 and Model 3. Over the intervals of 30 days, 90 days, 180 days, and 1 year, a continuous increase in CV was significantly associated with ACM across all models (P < 0.001). The highest tertile of CV (T3) exhibited a markedly elevated risk of ACM compared to T1, with the crude model showing HRs of 2.02 (95% CI 1.41–2.90; P < 0.001), 2.10 (95% CI 1.51–2.91; P < 0.001), 2.05 (95% CI 1.50–2.79; P < 0.001), and 1.99 (95% CI 1.48–2.67; P < 0.001) at 30 days, 90 days, 180 days, and 1 year, respectively. These significant associations persisted after adjustments in Model 1 and Model 2, and continued to show significance in Model 3 with HRs of 1.49 (95% CI 1.01–2.20; P = 0.046) at 30 days, 1.59 (95% CI 1.09–2.31; P = 0.016) at 90 days, 1.49 (95% CI 1.06–2.08; P = 0.021) at 180 days, and 1.46 (95% CI 1.06–2.01; P = 0.021) at 1 year, respectively. Furthermore, the implementation of RCS regression modeling disclosed that the mortality risk in the ICU, in-hospital, and at 30 days, 90 days, 180 days, and 1 year increased in a nonlinear fashion with rising CV (P_{non−linearity} = 0.002, P_{non−linearity} < 0.001, P_{non−linearity} = 0.002, P_{non−linearity} = 0.008, P_{non−linearity} = 0.007, and P_{non−linearity} = 0.021, respectively). This nonlinear relationship is illustrated in Fig. 3.

Fig. 3

RCS curve of GV and HR in patients with non-traumatic SAH A ICU stay; B in-hospital stay; C 30 days; D 90 days; E 180 days; F 1 year. RCS restricted cubic spline, GV glycemic variability, HR hazard ratio, SAH subarachnoid hemorrhage, CI confidence interval

Subgroup analysis

To evaluate the robustness of CV as a predictor of ACM, we performed stratified analyses across key demographic and clinical subgroups, including sex, age, hypertension, diabetes mellitus, myocardial infarction, congestive heart failure, MV, and antidiabetes therapy. The consistency of CV–ACM associations was assessed during ICU admission, hospitalization, and at 30 days, 90 days, 180 days, and 1-year intervals. As illustrated in Fig. 4, the association between elevated CV and increased ACM risk remained consistent across all subgroups (P for interaction > 0.05). The absence of significant interaction effects suggests that the predictive utility of CV for mortality is independent of sex, age, comorbidity burden, or critical care interventions. This stability in the CV–ACM relationship across diverse patient populations underscores its potential as a robust prognostic marker in critically ill patients with SAH.

Fig. 4

Subgroup analysis for the association between GV and ACM in patients with non-traumatic SAH. A ICU mortality; B in-hospital mortality; C 30-day mortality; D 90-day mortality; E 180-day mortality; F 1-year mortality. GV glycemic variability, ACM all-cause mortality, SAH subarachnoid hemorrhage

Discussion

This retrospective cohort study, leveraging data from the MIMIC-IV database, investigated the association between GV and ACM in critically ill patients with non-traumatic SAH. Our analyses demonstrated that elevated GV, quantified by CV, is an independent predictor of both short- and long-term ACM, even after comprehensive adjustment for demographic characteristics, clinical stabilization parameters, and comorbid conditions. RCS regression analysis revealed a ‘U’-shaped nonlinear relationship between GV and ACM in both short-term and long-term periods, indicating that beyond a certain threshold, higher GV levels are directly associated with an increased risk of mortality. Subgroup analysis revealed an absence of statistically significant interactions across different subgroups. These findings underscore the prognostic significance of GV in this high-risk population, suggesting its potential utility as a therapeutic target and risk stratification tool in neurocritical care settings. Unlike established aneurysmal SAH outcome models that primarily rely on clinical and imaging variables [17, 18], GV provides insights into metabolic instability as a prognostic marker. Future studies should investigate the integration of GV with traditional predictors to refine risk stratification and guide targeted therapeutic interventions.

In the ICU setting, daily artificial nutritional support, encompassing both enteral and parenteral nutrition, frequently leads to increased carbohydrate intake. This dietary pattern significantly elevates the risk of blood glucose disturbance [19]. Blood glucose abnormalities are highly prevalent among critically ill patients, with reported rates varying from 30 to 60%, contingent upon the specific care unit and diagnostic criteria utilized [20, 21]. Glycemic control is a critical component of managing critically ill patients. The implementation of intensive glycemic control in the ICU remains a subject of substantial controversy. Griesdale et al. demonstrated that intensive insulin therapy reduced mortality in critically ill patients in surgical ICUs [22]. A study conducted by Giakoumidakis and colleagues revealed that more intense blood glucose control resulted in significantly lower in-hospital mortality among cardiac surgery patients [23]. Although intensive insulin therapy has been associated with reduced infection rates and shorter ICU stays in neurosurgical patients [24, 25], it significantly increases hypoglycemia risk. Moreover, it has failed to demonstrate consistent mortality benefits and may even increase mortality in some critically ill populations [26,27,28]. Given these conflicting findings, the optimal glycemic management strategy for critically ill patients remains uncertain. The potential benefits of intensive insulin therapy—such as reduced mortality and improved outcomes in specific populations—must be balanced against the risks of hypoglycemia and potential harm. This complexity underscores the need for a more comprehensive approach beyond absolute blood glucose levels. In recent years, GV has emerged as a key indicator, capturing dynamic glucose fluctuations and providing a more nuanced understanding of glycemic control than traditional measures focused on average glucose levels.

The U-shaped relationship between GV and ACM, as revealed by RCS analysis, suggests that both excessively high and low GV levels may be detrimental. This nonlinear relationship aligns with previous studies that have highlighted the dual risks of hyperglycemia and hypoglycemia in critically ill patients. The increased mortality risk at higher GV levels may be attributed to the physiological stress caused by acute fluctuations in blood glucose, which can exacerbate endothelial dysfunction, oxidative stress, and inflammation, all of which are known to contribute to poor outcomes in SAH patients [29]. On the other hand, the increased risk at lower GV levels may reflect the potential harm of overly aggressive glycemic control, which can lead to hypoglycemia and its associated complications.

Our findings align with prior research demonstrating GV’s adverse effects in other critical care settings, such as traumatic brain injury and coronary artery disease, where higher GV correlates with increased mortality and worse outcomes [13, 14, 30]. The mechanisms underlying these associations are likely multifactorial, involving oxidative stress, endothelial dysfunction, and systemic inflammation, all of which are exacerbated by acute glucose fluctuations. In the context of SAH, these mechanisms may be particularly relevant, as the brain is highly sensitive to metabolic disturbances, and glucose variability could exacerbate secondary brain injury following the initial hemorrhage [31, 32]. In addition, our subgroup analyses, which did not show significant interactions across different groups, suggest that the association between GV and mortality is consistent across various demographic and clinical factors. This finding strengthens the generalizability of our results and indicates that GV could serve as a broadly applicable prognostic factor in critically ill patients. However, it is important to note that while GV is a significant predictor of mortality, other factors such as the severity of SAH, neurological status, and overall health condition must also be considered when determining a comprehensive treatment strategy. The use of CGM systems in the ICU could provide a more accurate and real-time assessment of GV, allowing for more precise glycemic management [33, 34]. The potential impact of CGM systems on improving clinical outcomes in patients with non-traumatic SAH warrants further investigation.

This study has several limitations that should be acknowledged. First, as a retrospective analysis, it is subject to the inherent biases and limitations of observational studies. Although we adjusted for multiple confounders, residual confounding (e.g., non-standardized measurement timing, detailed glycemic management strategies, or delayed cerebral ischemia) may persist. Real-world blood glucose sampling schedules inherent to ICU databases limit granular GV characterization, necessitating cautious interpretation of these findings. Second, the MIMIC-IV database, while comprehensive, is derived from a single center, which may limit the generalizability of our findings. Third, the definition of GV based on the CV may not capture all aspects of GV, and other measures such as MAGE could provide additional insights. Finally, our cohort exclusively comprised critically ill patients with non-traumatic SAH; thus, the applicability of these findings to other stroke subtypes or less severe cases requires further validation.

Despite these limitations, this study has several strengths. The large sample size of our cohort enhances the statistical power and reliability of our findings. In addition, the use of the MIMIC-IV database, a well-validated and comprehensive resource, ensures high-quality data and robust methodology. These strengths contribute to the overall credibility of our research.

Conclusion

This study demonstrates that elevated GV is an independent predictor of ACM in critically ill patients with non-traumatic SAH. The U-shaped relationship between GV and mortality underscores the importance of maintaining optimal glucose stability in this population. These findings support the incorporation of GV as a prognostic marker in SAH management and suggest that interventions targeting GV reduction may improve clinical outcomes. Future research should prioritize the development of targeted glycemic stabilization strategies and evaluate the utility of CGM systems in ICU settings to optimize glucose control and enhance patient outcomes.