Interpretable AI Model Successfully Predicts 28-Day Mortality Risk for ICU Patients with Sepsis and Respiratory Failure

Researchers develop an interpretable machine learning model to predict 28-day mortality for sepsis patients with respiratory failure using 20 clinical features.

By: AXL Media

Published: Mar 28, 2026, 5:22 AM EDT

Source: Information for this report was sourced from the Journal of Intensive Medicine

Bridging the Gap in Sepsis Prognosis

Sepsis remains a leading cause of death in intensive care units (ICUs) globally, with acute respiratory failure (ARF) standing as one of its most lethal complications. When these conditions intersect, patients often spiral into severe hypoxemia and multi-organ dysfunction within hours. Historically, accurately predicting which patients will survive the first month has been a significant clinical hurdle. Dr. Jian Liu and his team addressed this by creating a predictive model designed to function at the earliest stages of ICU admission, allowing for the rapid optimization of treatment strategies and the precise allocation of limited monitoring resources.

A "Training Plus External Validation" Framework

To ensure the model’s reliability across diverse healthcare environments, the researchers utilized two major global datasets. The model was initially developed and trained using the Medical Information Mart for Intensive Care (MIMIC-IV) database, which contains high-resolution data from thousands of ICU stays. Crucially, the team then performed an independent external validation using the eICU Collaborative Research Database (eICU-CRD). This rigorous dual-cohort design confirms that the model’s predictive power is not limited to a single hospital’s data but is applicable to real-world clinical settings internationally.

Harnessing XGBoost and Feature Selection

The researchers systematically compared seven different machine learning algorithms—including random forests and neural networks—finding that the XGBoost (Extreme Gradient Boosting) model delivered the highest level of accuracy. Using the Boruta feature selection algorithm and multicollinearity analysis, the team narrowed down hundreds of variables to 20 key predictive features. All 20 indicators, which cover oxygenation status, metabolic parameters, and organ function, are routinely collected during a patient’s first 24 hours in the ICU, making the model highly practical for bedside use.