Highlight
- The combination of Stress Hyperglycemia Ratio (SHR) and Glucose Variability (GV) serves as a superior predictor of 28-day mortality in sepsis compared to either metric alone.
- Mortality risk patterns are highly dependent on baseline metabolic status: High SHR/High GV is most lethal in NGR patients, while Low SHR/High GV carries the highest risk in Pre-DM patients.
- Machine learning models, specifically Random Forest and Logistic Regression, achieved an AUC of 0.776 in predicting clinical outcomes.
- Personalized glycemic management strategies are necessitated by the divergent impact of glucose fluctuations across different patient phenotypes.
Background: The Dual Threat of Acute Dysglycemia in Sepsis
Sepsis remains a leading cause of mortality in intensive care units (ICUs) worldwide, characterized by a dysregulated host response to infection that frequently triggers profound metabolic disturbances. Among these, stress-induced hyperglycemia is a hallmark of the acute illness phase, driven by the release of counter-regulatory hormones and proinflammatory cytokines. However, absolute glucose levels often fail to provide a complete picture, as they do not account for a patient’s baseline glycemic state or the degree of fluctuation experienced during the ICU stay.
Two critical metrics have emerged to address these gaps: the Stress Hyperglycemia Ratio (SHR), which adjusts acute glucose levels against chronic glycemic control (estimated by HbA1c), and Glucose Variability (GV), which captures the instability of blood sugar levels over time. While both have been independently linked to poor outcomes, their combined prognostic value—particularly when stratified by pre-existing metabolic conditions like prediabetes (Pre-DM) and diabetes mellitus (DM)—has remained largely unexplored. This study aimed to bridge this knowledge gap using large-scale real-world data and interpretable machine learning.
Study Design: A Deep Dive into the MIMIC-IV Cohort
This observational cohort study utilized data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, a comprehensive repository of de-identified electronic health records. Researchers screened thousands of sepsis patients, ultimately enrolling 4,838 individuals with a median age of 68 years. The cohort was stratified into three distinct groups based on their baseline glucose metabolic status: Normal Glucose Regulation (NGR), Prediabetes (Pre-DM), and Diabetes Mellitus (DM).
The primary endpoint was 28-day all-cause mortality, with ICU mortality serving as a secondary outcome. The researchers calculated SHR as the ratio of admission glucose to estimated average glucose derived from HbA1c. GV was determined based on the coefficient of variation of glucose measurements during the first 24 to 48 hours of ICU admission. To interpret the complex interactions between these variables, the team employed five machine learning (ML) algorithms, including XGBoost, Random Forest, and Logistic Regression, utilizing SHapley Additive explanations (SHAP) to ensure the models remained transparent and clinically interpretable.
Key Results: Divergent Risk Profiles Across Metabolic States
The study’s findings underscore the necessity of a nuanced approach to glycemic monitoring. Overall, 13.2% of patients died in the ICU, and 19.3% died within 28 days. However, the impact of SHR and GV varied dramatically across the metabolic subgroups.
The NGR Phenotype: The Danger of Extreme Fluctuation
In patients with normal baseline glucose regulation, the combination of high SHR (>1.23) and high GV (>28.56) resulted in the highest risk of death. These patients faced a Hazard Ratio (HR) of 2.06 (95% CI: 1.40-3.04) for 28-day mortality. This suggests that in previously healthy metabolic systems, the sudden onset of high stress-induced glucose coupled with erratic fluctuations is particularly poorly tolerated.
The Pre-DM Phenotype: Unexpected Vulnerability to Low SHR
Intriguingly, the Pre-DM group showed a different pattern. The highest mortality risk (HR = 2.45, 95% CI: 1.73-3.48) was observed in those with Low SHR (28.56). This finding suggests that for prediabetic patients, the inability to maintain a steady glucose level (high variability) is a more significant threat than the absolute peak of stress hyperglycemia itself.
The DM Phenotype: Chronic Adaptation and High SHR
For patients with established diabetes, the highest risk (HR = 1.46, 95% CI: 1.06-2.01) was found in the High SHR (>1.23) and Low GV (<28.56) group. Diabetic patients may have a higher physiological tolerance for glucose variability due to chronic exposure, but a significant spike relative to their already elevated baseline (High SHR) remains a potent predictor of mortality.
The Power of Interpretable Machine Learning
The integration of machine learning provided a robust framework for validating these clinical observations. Among the five models tested, Random Forest and Logistic Regression emerged as the top performers, both achieving an Area Under the Curve (AUC) of 0.776. The XGBoost model followed closely with an AUC of 0.746.
The use of SHAP values allowed the researchers to “open the black box” of the ML models, identifying SHR and GV as top-tier predictors alongside traditional markers like age and SOFA (Sequential Organ Failure Assessment) scores. This confirms that metabolic instability is not merely a bystander in sepsis but a core component of the disease’s pathophysiology.
Expert Commentary: Moving Toward Personalized Glycemic Targets
The results of this study challenge the “one-size-fits-all” approach to glycemic control in the ICU. For decades, the medical community has debated the optimal glucose target for critically ill patients. These findings suggest that the target should likely be dynamic, taking into account the patient’s prior metabolic history.
The biological plausibility of these findings lies in the concept of “glucose toxicity” versus “metabolic conditioning.” Diabetic patients may be conditioned to handle higher glucose levels, making the ratio (SHR) more important than the variation. Conversely, in NGR patients, the sudden loss of glucose homeostasis (high GV) may trigger oxidative stress and mitochondrial dysfunction more acutely. The study’s limitations include its retrospective nature and the reliance on the MIMIC-IV database, which may not capture all confounding variables such as nutritional support or specific insulin protocols used during the study period.
Conclusion: Summary and Future Directions
This observational cohort study demonstrates that the simultaneous assessment of SHR and GV provides a sophisticated tool for mortality risk stratification in sepsis. By identifying that different metabolic groups respond uniquely to stress hyperglycemia and glucose instability, the study paves the way for personalized glycemic management strategies. Clinicians should consider both the magnitude of the stress response (SHR) and the stability of the metabolic state (GV) when managing septic patients. Further prospective, randomized controlled trials are needed to determine if interventions specifically targeting these metrics can improve survival rates in the ICU.
References
Wang F, Guo Y, Jiao C, Zhao S, Sui L, Mao Z, Lu R, Hou R, Zhu X. Simultaneous assessment of stress hyperglycemia ratio and glucose variability to predict all-cause mortality in sepsis patients across different glucose metabolic states: an observational cohort study with interpretable machine learning approach. Int J Surg. 2026 Jan 1;112(1):1219-1232. doi: 10.1097/JS9.0000000000003525. Epub 2025 Sep 23. PMID: 40990680; PMCID: PMC12825658.