📄 medicine

Endothelial Activation and Stress Index as a Prognostic Marker and Machine Learning Predictor of Mortality in Mechanically Ventilated ICU Patients: A Multicenter Retrospective Cohort Study

This multicenter retrospective study demonstrates that the Endothelial Activation and Stress Index (EASIX) is an independent predictor of 28- and 90-day mortality in mechanically ventilated ICU patients and serves as a significant feature within an externally validated machine learning framework for risk stratification.

Original authors: Tianen Zhou, Qiaohua Hu, Juan Feng, Jingtao Xu, Guojun Chen

Published 2026-06-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tianen Zhou, Qiaohua Hu, Juan Feng, Jingtao Xu, Guojun Chen

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 human body as a bustling city. The endothelium is the city's plumbing system—the inner lining of all the blood vessels that keeps everything flowing smoothly and contained. When a patient gets very sick in the Intensive Care Unit (ICU) and needs a machine to breathe for them (mechanical ventilation), this "plumbing" often gets damaged. It starts leaking, clotting, and failing, which is a major reason why patients might not survive.

This paper is like a detective story where researchers tried to figure out if a specific "damage report" could predict who would survive and who wouldn't.

The "Damage Report" (EASIX)

The researchers focused on a score called EASIX (Endothelial Activation and Stress Index). Think of EASIX as a three-part weather report for the body's plumbing. It doesn't require expensive, rare tests; it just uses three numbers that doctors already check every day:

LDH: A measure of how much "debris" is in the blood from damaged cells (like broken bricks from a collapsing building).
Creatinine: A measure of how well the kidneys are filtering (like checking if the water treatment plant is working).
Platelets: A measure of the blood's clotting ability (like checking if the city's emergency repair crews are being used up too fast).

The researchers took these three numbers, did a little math to make them easier to read (called "log2-transformed"), and called it Log2EASIX.

The Investigation

The team looked at two huge digital archives of patient records:

The Main Archive (MIMIC-IV): Over 5,300 patients from a hospital in Boston.
The Test Archive (eICU): About 243 patients from hospitals across the US, used to see if the findings held up in a different group.

They wanted to answer two questions:

Does a higher "damage report" score mean a higher chance of dying within 28 days or 90 days?
Can we use this score, combined with other data, to build a computer program (Machine Learning) that predicts death better than current methods?

What They Found

1. The Score is a Crystal Ball
The results were clear: The higher the damage report (Log2EASIX), the higher the risk of death.

It wasn't a "pass/fail" switch. Instead, it was like a volume knob. As the score went up, the risk of dying went up steadily. There was no "safe zone" where the score stopped mattering; every little increase in damage added more risk.
Even after accounting for how sick the patient was, their age, and their other diseases, the score still predicted death independently. It was like finding a new, independent witness who confirmed the danger, even when other witnesses were already speaking.

2. The Computer Predictor
The researchers built a "smart" computer model (using a method called CatBoost) to predict who would die within 28 days. They fed it 19 different pieces of information, including the damage report (Log2EASIX), age, and severity scores.

The computer learned that the damage report (Log2EASIX) was the third most important clue in predicting death, right behind the patient's age and a standard severity score called APS III.
The computer model worked well on the main group of patients and held up reasonably well when tested on the second group of patients from different hospitals.

3. The "Net Benefit"
They used a tool called Decision Curve Analysis to see if using this model would actually help doctors make better decisions. Imagine a scale: on one side is "treat everyone," and on the other is "treat no one." The study showed that using this model tipped the scale in the right direction, offering a real benefit over just guessing or treating everyone blindly.

The Bottom Line

This study suggests that the EASIX score is a powerful, simple tool. It's like a "stress gauge" for the body's blood vessels that can be calculated instantly from routine blood tests.

For the patient: A high score means the body's plumbing is under severe stress, and the risk of not surviving the next month or three months is significantly higher.
For the doctor: It provides an extra layer of insight. Even if a patient looks "okay" on standard severity scores, a high EASIX score might warn that their internal plumbing is failing, helping doctors understand the true risk.

The researchers concluded that this score is a reliable, independent predictor of mortality for patients on ventilators and could be a useful tool for sorting patients by risk right at the bedside. However, they noted that because this was a look-back study (using old data), it proves a connection but doesn't yet prove that changing the score will save lives—that would require future tests.

Technical Summary: Endothelial Activation and Stress Index as a Prognostic Marker and Machine Learning Predictor in Mechanically Ventilated ICU Patients

Problem Statement
Mechanical ventilation (MV) is a critical intervention for acute respiratory failure, yet mortality rates among mechanically ventilated intensive care unit (ICU) patients remain substantial (25–45%). While endothelial dysfunction is a recognized pathophysiological mechanism linking systemic inflammation to multi-organ failure, routine clinical assessment of endothelial status is limited due to the cost and unavailability of specific biomarkers (e.g., angiopoietin-2). The Endothelial Activation and Stress Index (EASIX), calculated from routine laboratory values (lactate dehydrogenase, serum creatinine, and platelet count), has shown prognostic value in hematological and cardiological populations but lacks systematic evaluation in mechanically ventilated cohorts. Furthermore, existing prognostic studies often rely on traditional regression models that may fail to capture complex, non-linear interactions among predictors, and many machine learning (ML) studies lack external validation and model transparency.

Methodology
This study employed a multicenter retrospective cohort design utilizing two publicly available databases:

Primary Cohort: MIMIC-IV (version 3.1), containing data from 5,300 mechanically ventilated adult patients admitted to the ICU for the first time between 2008 and 2019.
External Validation Cohort: eICU Collaborative Research Database (version 2.0), containing 243 patients from over 200 US hospitals (2014–2015).

Inclusion/Exclusion: Patients were included if they were ≥18 years old, received invasive MV for ≥24 hours, had an ICU stay ≥2 days, and possessed complete data for LDH, platelets, and creatinine.

Variable Construction:

Exposure: EASIX was calculated as $[LDH \times Creatinine] / Platelets$ . Due to right-skewness, the log2-transformed value (Log2EASIX) was used for all analyses.
Outcomes: Primary outcome was 28-day all-cause mortality; secondary was 90-day all-cause mortality.
Covariates: A comprehensive set including demographics, vital signs, laboratory parameters, illness severity scores (SOFA, SAPS II, APS III), and comorbidities.

Statistical and Machine Learning Approach:

Regression Analysis: Cox proportional-hazards regression with restricted cubic splines (RCS) evaluated the dose-response relationship between Log2EASIX and mortality across four sequentially adjusted models.
Feature Selection: A consensus feature set was derived using Boruta (ensemble wrapper) and LASSO regression (Cox loss function) on the training set (70% of MIMIC-IV).
Model Training: Fifteen ML classifiers were trained on the consensus feature set (19 variables). Model selection was based on benchmark cross-validated AUC-ROC to penalize overfitting.
Interpretability & Utility: The optimal model was interpreted using SHapley Additive exPlanations (SHAP) for feature importance and directional effects. Decision Curve Analysis (DCA) assessed net clinical benefit against "treat-all" and "treat-none" strategies.

Key Results

Independent Association: In the fully adjusted model, each one-unit increment in Log2EASIX was associated with a 25% higher hazard of 28-day mortality (HR 1.25, 95% CI 1.19–1.32) and a 21% higher hazard of 90-day mortality (HR 1.21, 95% CI 1.16–1.27).
Dose-Response: RCS analysis revealed a predominantly log-linear relationship between Log2EASIX and mortality risk, with no evidence of a threshold effect or plateau.
Subgroup Analysis: The association was consistent across most subgroups. After Bonferroni correction, statistically robust effect modifications for 90-day mortality were identified only for GCS score, Charlson Comorbidity Index, and malignant cancer status.
Machine Learning Performance:
- Optimal Model: CatBoost achieved the highest discriminative performance with a cross-validated AUC of 0.758.
- External Validation: The model maintained performance on the independent eICU cohort (AUC 0.745).
- Feature Importance: SHAP analysis ranked Log2EASIX as the third most important predictor (mean |SHAP| = 0.0497), following APS III score and age.
- Clinical Utility: DCA confirmed that the CatBoost model provided a net clinical benefit over standard strategies across a broad range of threshold probabilities in both internal and external validations.

Key Contributions

Prognostic Validation: This study establishes Log2EASIX as an independent predictor of short- and medium-term mortality in mechanically ventilated ICU patients, extending its applicability beyond hematological and cardiological contexts.
Continuous Risk Gradient: The research demonstrates that the relationship between EASIX and mortality is log-linear, supporting the use of EASIX as a continuous risk gradient rather than a binary cutoff.
ML Framework with Transparency: The study develops an externally validated ML model where EASIX is a top-tier feature, utilizing SHAP to ensure model interpretability and DCA to validate clinical utility, addressing common limitations in critical care ML studies.
Practicality: The index relies solely on three routinely available laboratory values, offering a cost-effective tool for bedside risk stratification.

Significance and Claims
The authors claim that Log2EASIX provides meaningful prognostic information that is not fully subsumed by established illness-severity scores (SOFA, SAPS II, APS III). By integrating EASIX into a machine learning framework, the study supports its potential role as a practical bedside tool for early risk stratification in high-mortality populations. The findings suggest that endothelial stress, captured by this simple composite index, is a critical determinant of outcomes in mechanically ventilated patients. The study concludes that these results warrant further prospective investigation to confirm EASIX's utility in guiding clinical decision-making, such as monitoring intensity and goals-of-care discussions.

The "Damage Report" (EASIX)

The Investigation

What They Found

The Bottom Line

Technical Summary: Endothelial Activation and Stress Index as a Prognostic Marker and Machine Learning Predictor in Mechanically Ventilated ICU Patients

More like this