What is the ideal glucose range for a patient with sepsis in the ICU? A retrospective analysis of MIMIC-IV

Tristan Struja, Lasse Hansen, João Matos, Josep Gómez, A. Pardo, Ismini Lourentzou, Nima S. Hejazi, Leo Anthony Celi, Andre Kurepa WaschkaBMJ 2026 · DOI

TL;DR

This study uses advanced causal inference techniques on a large ICU dataset (MIMIC-IV) to investigate the optimal blood glucose range for patients with sepsis. The findings suggest a U-shaped relationship between glucose levels and mortality, with an optimal range of 160-190 mg/dL, and importantly, tighter glucose control did not improve outcomes and increased the risk of hypoglycemia.

What Problem Does This Paper Address?

Managing blood glucose levels in critically ill patients, particularly those with sepsis in the Intensive Care Unit (ICU), is a long-standing clinical challenge. While tight glucose control (aiming for lower blood sugar) was once a popular strategy, recent large-scale randomized controlled trials (RCTs) have yielded conflicting results. Some studies suggested benefits, while others found no improvement and even increased mortality, often linked to a higher incidence of hypoglycemia. This ambiguity leaves clinicians uncertain about the ideal glucose target, especially when considering individual patient characteristics and the potential for confounding factors in observational data. This paper aims to clarify this by using sophisticated methods to emulate a trial and analyze real-world data, seeking an optimal glucose range independent of confounding variables.

Methods

The researchers employed a two-pronged analytical approach:

  1. Targeted Maximum Likelihood Estimation (TMLE): This is a powerful causal inference framework that uses machine learning to estimate average treatment effects. In this study, TMLE was used to emulate a "targeted trial" by shifting the observed glucose distributions to predefined means (100, 130, 160, 190, and 220 mg/dL) and estimating the resulting impact on outcomes. TMLE is known for its "doubly robust" nature, meaning it provides unbiased estimates if either the outcome regression or the exposure mechanism is correctly specified.
  2. Joint Longitudinal-Survival Modeling: This approach combines longitudinal data (e.g., hourly SOFA scores, insulin doses) with time-to-event data (e.g., death or discharge). It allows for the simultaneous modeling of how patient conditions evolve over time and their eventual outcome, capturing the dynamic interplay between glucose levels, organ dysfunction, and mortality.

These methods were chosen to leverage the high-resolution data available in MIMIC-IV and to overcome limitations of traditional observational studies by attempting to establish causal relationships.

Datasets

The study utilized the Medical Information Mart for Intensive Care (MIMIC-IV) database. This is a large, publicly available, de-identified ICU database containing comprehensive physiological data, laboratory results, medications, diagnoses, and demographics from patients admitted to Beth Israel Deaconess Medical Center between 2008 and 2019. The researchers extracted data from approximately 70,000 de-identified ICU records.

Key Insights and Ideas

  • The "Sweet Spot" for Glucose: The most striking finding is the suggestion of a U-shaped relationship between blood glucose and mortality. This implies that both very low and very high glucose levels are detrimental, with an optimal range existing in the middle.
  • Liberal Glucose Control May Be Best: Contrary to some older beliefs, the study indicates that aiming for a liberal glucose range (around 160-190 mg/dL) might be more beneficial than aggressive tight control.
  • Hypoglycemia is a Major Risk: Aggressively lowering glucose levels significantly increased the frequency of hypoglycemic events (<80 mg/dL), which are themselves associated with adverse outcomes. This risk appears to outweigh any potential benefits of tighter control.
  • Diabetes Status Matters, But Not for Lowering Targets: While diabetes status influenced baseline glucose levels and the frequency of hypoglycemia, the study did not find specific patient subgroups that would benefit from lower glucose targets.
  • Stress Hyperglycemia as Adaptive: The paper posits that stress hyperglycemia in sepsis might be a partly adaptive metabolic response, particularly in non-diabetic patients, which could explain why forcing glucose levels lower doesn't improve outcomes.

Key Results

  • Mortality: The study found that mortality was lowest (around 17%) with mean glucose levels between 130 and 160 mg/dL. Mortality increased at both lower (21% at 100 mg/dL) and higher (26% at 220 mg/dL) glucose levels, though the overall trend for mortality across glucose strata was not statistically significant (p-for-trend 0.26).
  • Hypoglycemia: The incidence of hypoglycemic events (<80 mg/dL) increased dramatically with tighter glucose control, rising from 16% at 220 mg/dL to 77% at 100 mg/dL (p-for-trend 0.01). Severe hypoglycemia (<50 mg/dL) also increased significantly with lower glucose targets.
  • Secondary Outcomes: There was a trend towards more 28-day hospital-free and ICU-free days with lower glucose target ranges, but illness severity (SOFA scores) did not change with different glucose targets.

Limitations and Future Work

  • Single-Center Design: The study was conducted at a single academic hospital in Boston, USA. This limits generalizability, as the patient population may not reflect the demographics and clinical characteristics of broader ICU populations. The cohort was older, more comorbid, and had a higher prevalence of diabetes than the general population.
  • Unmeasured Confounding: As with all observational studies, the causal inference relies on the assumption that all relevant confounders were measured and appropriately adjusted for. The authors acknowledge that this assumption can never be fully verified.
  • Missing Data: While sophisticated methods were used, missing data in ICU datasets is complex and can introduce bias.
  • Future Work: The authors suggest that the adoption of target trial emulations and causal inference frameworks in high-resolution healthcare data should become the norm for planning prospective studies, potentially reducing the need for costly and sometimes harmful RCTs. They also highlight the need for further research to understand the specific mechanisms behind the observed relationships.

Summary updated

Abstract

Our data suggest a U-shaped association of glucose and mortality with an optimal average glucose between 160 and 190 mg/dL. These results confirm current guideline recommendations. Together with recent results from randomised controlled trials, intensivists should aim for a liberal glucose range in most patients.
BibTeX
@inproceedings{struja2026what,
  author = {Tristan Struja and Lasse Hansen and João Matos and Josep Gómez and A. Pardo and Ismini Lourentzou and Nima S. Hejazi and Leo Anthony Celi and Andre Kurepa Waschka},
  title = {{What is the ideal glucose range for a patient with sepsis in the ICU? A retrospective analysis of MIMIC-IV}},
  booktitle = {BMJ},
  year = {2026},
  doi = {10.1136/bmjopen-2025-104916},
}
Figure