Sepsis is one of the most common and lethal syndromes encountered in intensive care units (ICUs), and acute respiratory failure (ARF) represents one of its most critical complications. Once respiratory failure develops, patients often experience severe hypoxemia and multiple organ dysfunction within a short period of time, resulting in a markedly increased risk of death. Despite advances in critical care, accurately assessing short-term prognosis early after ICU admission remains a major challenge in clinical practice.
In a recent study, a research comprising Dr. Jian Liu from the Gansu Provincial Maternity and Child Health Hospital (Gansu Provincial Central Hospital), China, Engineer Zi Yang from The First Hospital of Lanzhou University, China, Dr. Hong Guo at the Gansu Provincial Maternity and Child Health Hospital (Gansu Provincial Central Hospital), China, among other researchers developed and validated a machine learning model to predict 28-day mortality in patients with sepsis complicated by ARF. The results of this study were published online in the Journal of Intensive Medicine on January 10, 2026.
Speaking on the study, Dr. Liu says, "The model was designed to leverage clinical information available at the earliest stage of ICU admission, enabling clinicians to identify high-risk patients promptly and thereby optimize treatment strategies and the allocation of monitoring resources."
The Medical Information Mart for Intensive Care IV (MIMIC-IV, version 3.1) database was used as the development and training cohort, including adult ICU patients who met the diagnostic criteria for both sepsis and ARF. To evaluate the model's applicability across different hospitals and patient populations, an independent external validation was performed using data from the eICU Collaborative Research Database (eICU-CRD, version 2.0). This combined 'training plus external validation' design enhances the relevance of the findings to real-world clinical settings.
During variable selection, candidate predictors were first identified based on the international sepsis-related guidelines and expert clinical consensus to ensure strong clinical relevance. The Boruta feature selection algorithm, together with multicollinearity analysis, was then applied to identify a final set of 20 key predictive features. All selected variables were routinely obtainable within the first 24 hours of ICU admission and reflected multiple clinical dimensions, including oxygenation status, organ function, metabolic parameters, and disease severity.
Seven machine learning algorithms, including logistic regression, random forests, gradient-boosting, and neural networks, were systematically compared. Among them, the XGBoost model demonstrated the best overall performance. In the training cohort, the model showed strong discrimination for predicting 28-day mortality, and its performance remained stable in the independent external validation cohort, indicating good generalizability. Unlike traditional 'black-box' prediction models, this study placed particular emphasis on interpretability. The researchers applied SHapley Additive exPlanations (SHAP) to quantify the contribution of individual clinical variables to mortality risk prediction.
"Our analysis highlighted the importance of oxygenation indices, serum albumin levels, liver function-related indicators, and disease severity scores in short-term prognosis. This transparent interpretability framework may facilitate clinician understanding and promote the use of the model as a decision-support tool rather than a replacement for clinical judgment," explains Engineer Zi Yang and Dr. Hong Guo.
According to the research team, the model may be further integrated into bedside or web-based risk assessment tools to support early risk stratification in patients with sepsis complicated by ARF. Overall, the study demonstrates the potential of interpretable machine learning approaches in critical care medicine and provides a new technical pathway for individualized management of high-risk patients with sepsis.
