Development and external validation of a machine learning model for predicting the 28-day mortality risk in patients with sepsis complicated by acute respiratory failure in the ICU

Sepsis complicated by acute respiratory failure (ARF) is a common and severe condition among patients admitted to the intensive care unit (ICU), carrying a high mortality risk. Accurate prediction of short-term outcomes is crucial for optimizing clinical decisions and treatment strategies. Therefore, we aimed to develope and validated an interpretable machine learning (ML) model to predict the 28-day mortality risk in ICU patients with sepsis complicated by ARF. This retrospective study included ICU patients with sepsis complicated by ARF from the Medical Information Mart for Intensive Care IV (MIMIC-IV, v3.1) database as the training cohort, and patients from the eICU-CRD (v2.0) database as the external validation cohort. Candidate predictors were initially identified based on clinical guidelines and expert consensus, and the Boruta algorithm was applied to determine the optimal feature set. Seven ML algorithms, namely random forest, XGBoost, logistic regression, AdaBoost, gradient boosting, CatBoost, and neural network, were compared. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. To enhance interpretability, feature importance was assessed using SHapley Additive exPlanations (SHAP) analysis. These results were integrated to construct a practical prognostic prediction platform. The training cohort (2975 deaths) comprised 12,597 ICU patients with sepsis complicated by ARF from the MIMIC-IV (v3.1) database, while 890 patients from the eICU-CRD (v2.0) database (102 deaths) were included for external validation. Among the ML models, XGBoost achieved the best performance in the training cohort (AUC: 0.812; accuracy: 0.772; sensitivity: 0.605; specificity: 0.823). In the external validation cohort, XGBoost demonstrated good generalizability (AUC: 0.714). SHAP analysis identified PaO₂, alanine aminotransferase, albumin, age, Acute Physiology and Chronic Health Evaluation (APACHE) III, lactate, urine output, and respiratory rate as the most influential predictors of 28-day mortality. Accordingly, a short-term mortality prediction platform was developed. We successfully developed an efficient, interpretable predictive model based on the XGBoost algorithm, accurately predicting 28-day mortality risk for ICU patients with sepsis complicated by ARF. Demonstrating stable performance and strong generalizability, this model holds promise as a clinical decision-support tool for the early identification of high-risk patients and optimization of personalized treatments.