📄 medicine

An Interpretable EHR-Derived Lactate–Hemodynamic Framework for Mortality Risk Stratification in ICU Sepsis: A MIMIC-IV Derivation and eICU Validation Study

This study demonstrates that an interpretable framework classifying ICU sepsis patients into four hypoperfusion phenotypes based on lactate, mean arterial pressure, and vasopressor use effectively stratifies mortality risk, with the combined hypoperfusion phenotype showing the strongest and most reproducible association with adverse outcomes across both MIMIC-IV and eICU cohorts.

Original authors: Xinlei He, Mengyuan Shen, Yunhui Ni, Jianfei Sun, Weiwei Chen, Xiaoxiang Gao, Haibo Qiu

Published 2026-07-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinlei He, Mengyuan Shen, Yunhui Ni, Jianfei Sun, Weiwei Chen, Xiaoxiang Gao, Haibo Qiu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Intensive Care Unit (ICU) as a high-stakes control room where doctors are trying to keep a patient's internal engine running smoothly. When a patient gets sepsis—a dangerous, body-wide infection—the engine starts sputtering. The big question is: How do we know which patients are about to crash completely?

For a long time, doctors have looked at two main dashboard lights to guess the answer:

The Lactate Gauge: A chemical in the blood that spikes when cells aren't getting enough oxygen (like a "low fuel" light).
The Pressure Gauge: The blood pressure and whether the patient needs strong drugs (vasopressors) to keep the heart pumping hard enough.

Some doctors thought, "If the pressure is low, that's the big problem." Others thought, "No, if the lactate is high, that's the real danger." This study, which looked at data from over 30,000 patients in two massive digital hospital databases (MIMIC-IV and eICU), decided to test a new idea: What if we look at the dashboard lights together?

The Four "Engine Status" Groups

The researchers sorted the patients into four groups based on what their dashboard looked like in the first 24 hours:

The "All Good" Group (Reference): Normal pressure, normal lactate. (About 6% of patients in the first group).
The "High Lactate" Group: Normal pressure, but the lactate light is flashing. (About 2% of patients).
The "Low Pressure" Group: Low pressure or on pressure-boosting drugs, but normal lactate. (About 75% of patients in the first group).
The "Double Trouble" Group (Combined Hypoperfusion): Both the pressure is low and the lactate is high. (About 18% of patients in the first group).

The Big Discovery: "Double Trouble" is the Real Danger

The study found that while any warning light is bad, the "Double Trouble" group was in the most serious shape.

Think of it like a car engine. If your car is making a weird noise (high lactate) but the speedometer is fine, it's concerning. If your speedometer is dropping (low pressure) but the engine sounds okay, that's also worrying. But if both the engine is screaming and the speedometer is plummeting? That's when you know the car is about to stop running entirely.

In the first group of patients (the "derivation" group), the death rates climbed a ladder as the problems got worse:

All Good: ~6% died in the hospital.
High Lactate only: ~15% died.
Low Pressure only: ~17% died.
Double Trouble: 30% died.

The researchers then tested this idea on a completely different group of patients (the "validation" group) to see if it was just a fluke. The result? The ladder held up. The "Double Trouble" group still had the highest death rate, with 40% of them dying in the hospital.

What This Rules Out

Here is where it gets interesting. The study explicitly tested whether "Low Pressure" alone was a reliable warning sign across different hospitals.

In the first group, low pressure alone was linked to higher death rates.
However, when they checked the second group, the "Low Pressure only" group lost its statistical significance. This means that in the second group, having low pressure without high lactate didn't reliably predict death after adjusting for other factors.

The paper argues against the idea that low pressure alone is a consistent, standalone predictor of death across all different hospital settings. It suggests that low pressure might be a "noisy" signal that changes meaning depending on the context, whereas the "Double Trouble" signal is loud and clear everywhere.

How Sure Are We?

The authors are very confident about the "Double Trouble" finding. They didn't just guess; they ran complex math models that adjusted for age, sex, and how sick the patient was overall (using a score called SOFA). Even after all that math, the "Double Trouble" group remained the one most strongly linked to death.

In the first group, patients with "Double Trouble" were 3.16 times more likely to die in the hospital than the "All Good" group.
In the second group, they were 3.15 times more likely to die.

This consistency across two huge, different databases makes the finding robust.

However, the authors are careful to say they haven't "solved" sepsis. They didn't prove that causing low pressure and high lactate makes people die (because this was a look-back study, not a new experiment). They also noted that the "Low Pressure" data in the second database was recorded differently (using a simple "yes/no" flag for drugs rather than detailed records), which might be why that specific group's results were shaky.

The Takeaway

The paper suggests that for doctors standing at the bedside, looking at both the pressure and the lactate together gives a much clearer picture of who is in immediate danger than looking at just one. If a patient has both signs, they are in the "Double Trouble" zone and need the most urgent attention.

It's a simple, easy-to-read framework that doesn't require complex computers—just the standard tools doctors already have. While the study doesn't claim this is a magic cure, it offers a reliable way to sort patients by risk, helping doctors decide who needs to be watched most closely. As the authors note, this is a "pragmatic bridge" between what's happening inside the body and what doctors can see on the monitor.

Technical Summary: An Interpretable EHR-Derived Lactate–Hemodynamic Framework for Mortality Risk Stratification in ICU Sepsis

Problem Statement
Despite advances in early recognition and hemodynamic resuscitation, sepsis remains a leading cause of death in critical care. Current clinical practice relies on individual bedside indicators—blood lactate, mean arterial pressure (MAP), and vasopressor use—to assess perfusion. However, interpreting these signals in isolation may oversimplify the pathophysiology of sepsis, as metabolic and hemodynamic abnormalities represent distinct but complementary dimensions of circulatory and cellular dysfunction. While the Sepsis-3 definition of septic shock (requiring vasopressor-dependent hypotension and elevated lactate despite fluid resuscitation) identifies a high-risk group, its formal application is often impractical at the bedside due to the difficulty of confirming adequate fluid resuscitation and persistent vasopressor requirements in observational datasets. Furthermore, the prognostic value of patients with isolated metabolic or isolated hemodynamic abnormalities (those falling outside the strict Sepsis-3 septic shock category) remains under-characterized and lacks external validation across independent cohorts. There is a need for a simplified, interpretable framework based on routinely available Electronic Health Record (EHR) variables that can stratify mortality risk without complex model infrastructure.

Methodology
The study employed a retrospective derivation and external validation design using two large, publicly available critical care databases:

Derivation Cohort: MIMIC-IV database ( $n = 20,076$ adults with sepsis).
Validation Cohort: eICU Collaborative Research Database ( $n = 10,756$ adults with sepsis).

Phenotype Definition:
Patients were classified into four mutually exclusive "early hypoperfusion phenotypes" based on the worst or most clinically relevant values recorded within the first 24 hours of ICU admission:

Reference: No hyperlactatemia ( $<2$ mmol/L) and no hemodynamic hypoperfusion.
Isolated Hyperlactatemia: Lactate $\ge 2$ mmol/L without hemodynamic hypoperfusion.
Hemodynamic Hypoperfusion: MAP $< 65$ mm Hg and/or vasopressor use, without hyperlactatemia.
Combined Hypoperfusion: Coexistence of hyperlactatemia and hemodynamic hypoperfusion.

Note on Vasopressors: In MIMIC-IV, vasopressor use was identified via medication records. In eICU, it was identified via a treatment-category flag ("vasopressors"), which captured a smaller proportion of patients. To address this heterogeneity, a planned sensitivity analysis in eICU redefined hemodynamic hypoperfusion using MAP $< 65$ mm Hg alone, excluding the vasopressor flag.

Statistical Analysis:
The primary outcome was hospital mortality; secondary outcomes included ICU mortality and 28-day mortality (derivation only). Multivariable logistic regression models were constructed to evaluate the association between phenotypes and mortality, adjusting for age, sex, SOFA score, creatinine, and pH.

Key Results

Phenotype Distribution: The distribution of phenotypes varied between cohorts, largely due to differences in vasopressor recording (64.0% in MIMIC vs. 5.4% in eICU). However, the combined hypoperfusion phenotype consistently exhibited the most severe baseline clinical profile (higher SOFA scores, higher lactate, lower MAP, greater fluid balance).
Mortality Gradients: In both cohorts, crude mortality rates increased stepwise across phenotypes. The combined hypoperfusion group consistently demonstrated the highest mortality rates across all endpoints (e.g., in MIMIC: 30.3% hospital mortality; in eICU: 40.1% hospital mortality).
Independent Associations (Derivation Cohort): After full adjustment, all non-reference phenotypes were associated with increased mortality.
- Combined Hypoperfusion: Showed the strongest and most consistent association with ICU mortality (OR 7.02) and hospital mortality (OR 3.16).
- Isolated Hyperlactatemia: Remained independently associated with all mortality endpoints, showing the largest effect for 28-day mortality (OR 2.45).
- Hemodynamic Hypoperfusion: Associated with increased mortality in the derivation cohort.
External Validation (eICU Cohort):
- Combined Hypoperfusion: Retained a robust, independent association with ICU mortality (OR 3.11) and hospital mortality (OR 3.15).
- Isolated Hyperlactatemia: Remained significantly associated with mortality.
- Hemodynamic Hypoperfusion: Lost statistical significance after full adjustment for ICU and hospital mortality (OR 1.05 and 1.11, respectively), suggesting its prognostic signal is less stable across cohorts when vasopressor data is limited.
Sensitivity Analysis: When hemodynamic hypoperfusion in eICU was defined solely by MAP $< 65$ mm Hg, the results remained essentially unchanged, confirming that the robust signal of the combined phenotype was not dependent on the specific operationalization of vasopressor exposure.

Key Contributions

Simplified Framework: The study proposes and validates a bedside framework using only lactate and hemodynamic status (MAP/vasopressors) to stratify risk, bypassing the need for formal confirmation of fluid adequacy required by strict Sepsis-3 septic shock definitions.
Phenotypic Differentiation: It demonstrates that while isolated metabolic and isolated hemodynamic abnormalities carry risk, the combined phenotype identifies a distinct, high-risk subgroup with the most reproducible mortality gradient across independent datasets.
Reproducibility: The study highlights that the prognostic value of the combined phenotype is robust across different data sources and recording granularities, whereas the prognostic stability of isolated hemodynamic hypoperfusion is more variable.

Significance and Claims
The authors claim that this interpretable, EHR-derived framework successfully reproduces the mortality gradients associated with Sepsis-3 septic shock without requiring the confirmatory steps that are often impractical in real-time clinical settings or retrospective data analysis. The findings suggest that metabolic and hemodynamic abnormalities are complementary rather than interchangeable; their coexistence identifies patients with more severe physiologic derangement and a markedly higher risk of adverse outcomes.

The study concludes that identifying combined hypoperfusion at ICU admission can support early risk stratification and clinical decision-making, potentially flagging patients who warrant closer reassessment of perfusion status and organ dysfunction. The authors emphasize that this framework serves as a pragmatic bridge between sepsis pathophysiology and bedside decision support, though they note that prospective validation is required before clinical implementation. They explicitly state that the combined phenotype should be interpreted as "Sepsis-3–aligned" rather than equivalent to a formal Sepsis-3 septic shock diagnosis due to the retrospective nature of the data and the inability to fully verify fluid resuscitation adequacy.