A multicenter machine learning model for predicting ICU mortality in mechanically ventilated patients: development and external validation.

Abstract

Background: Accurate early prediction of mortality in mechanically ventilated intensive care unit (ICU) patients remains challenging due to disease heterogeneity and dynamic clinical trajectories. Traditional severity scoring systems have limited flexibility and generalizability. This study aimed to develop and externally validate a machine learning-based model using routinely available clinical variables to predict ICU mortality in mechanically ventilated patients.

Methods: A retrospective multicenter cohort study was conducted using the MIMIC database as the development cohort, with external validation performed in the eICU Collaborative Research Database and an independent institutional cohort. Adult ICU patients receiving invasive mechanical ventilation were included. A two-step feature selection strategy combining least absolute shrinkage and selection operator (LASSO) regression and recursive feature elimination was applied to identify robust predictors. Six machine learning algorithms were compared, and a random forest model was selected. Model performance was evaluated using discrimination, calibration, decision curve analysis, time-dependent area under the curve (AUC), and concordance index (C-index). Model interpretability was assessed using SHAP analysis.

Results: A total of 21 predictors were selected for model development. The random forest model demonstrated superior overall performance compared with alternative algorithms. In external validation, the model showed favorable early discrimination, with near-perfect AUC values during the first 48 hours in the local cohort and consistently acceptable performance in the eICU cohort. Predictive performance attenuated over time, particularly in the local cohort, whereas temporal stability was observed in the multicenter dataset. Calibration and decision curve analyses supported the reliability and clinical utility of the model. Key predictors included age, markers of respiratory dysfunction, metabolic stress, organ failure, and treatment intensity.

Conclusions: The proposed random forest model provides accurate and interpretable admission-time risk stratification for ICU mortality in mechanically ventilated patients. We recommend its use as an admission triage and screening tool for identifying high-risk patients within the first 48 hours of ICU admission. While predictive performance declined over time, the model demonstrated robust early risk stratification and generalizability across independent cohorts, highlighting its potential value for early clinical decision support in critical care settings.

Clinical trial number: Not applicable. This study was an observational retrospective cohort study and was not registered as a clinical trial.