Stacking machine learning model for risk stratification of acute respiratory distress syndrome after traumatic brain injury: a multicenter retrospective study.

BACKGROUND: Acute respiratory distress syndrome (ARDS) is a severe complication after traumatic brain injury (TBI), and early risk stratification may facilitate timely preventive interventions and improve clinical outcomes. METHODS: Model development used the MIMIC-IV v3.1 database with external validation in a cohort from the First Affiliated Hospital of Xinjiang Medical University. Candidate variables recorded within 24 hours of ICU admission were sequentially screened by univariable analysis, Boruta, and ElasticNet to identify predictors for model training. Eleven machine-learning algorithms were trained and integrated into a stacking ensemble optimized through hyperparameter tuning and five-fold cross-validation. Performance was evaluated using discrimination, calibration, decision-analytic metrics, and standard classification measures, with SHAP used for model interpretation. RESULTS: The development cohort included 985 patients (280 ARDS) and the external validation cohort 184 patients (67 ARDS). The stacking ensemble demonstrated stable discrimination across training, internal test, and external validation cohorts (AUC 0.948, 0.926, and 0.912; AP 0.950, 0.828, and 0.863), with good calibration and favorable decision-analytic performance. SHAP analysis identified global severity, hemodynamic instability, neurologic dysfunction, and metabolic disturbances as major contributors. CONCLUSION: The stacking model showed good performance and interpretability, suggesting potential utility for ARDS risk stratification in critically ill TBI patients.