Early Prediction of Standing at Discharge in Moderate-to-Severe TBI: A Clinical Machine Learning Model Integrating Modifiable and Nonmodifiable Factors.

Abstract

To develop and internally validate a machine learning model to predict favorable standing ability at hospital discharge in patients with moderate-to-severe traumatic brain injury (TBI), incorporating both modifiable and nonmodifiable clinical factors. Retrospective cohort study. A tertiary academic medical center in Taiwan. A total of 248 adults with moderate-to-severe TBIs admitted between 2019 and 2024 who received standard acute care and had complete discharge functional outcome data. Not applicable. Favorable standing ability at discharge, defined as a score ≥6 on the Modified ICU Mobility Scale. Predictor variables included age, Glasgow Coma Scale score, Injury Severity Score, Charlson Comorbidity Index, alanine aminotransferase level, standardized education scale (nonmodifiable), intubation duration, early mobilization, trauma activation, and selected micronutrient supplementation (modifiable). Four ML models-logistic regression, extreme gradient boosting (XGBoost), random forest, and support vector machine-were trained using an 80/20 data split. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), Brier score, and decision curve analysis (DCA). The XGBoost model achieved the greatest discrimination (AUC = 0.85), good calibration (Brier score = 0.16), and an accuracy of 78%, followed by the logistic regression model (AUC = 0.82; Brier score = 0.16; accuracy = 80%). The most influential predictors were age, intubation duration, and early mobilization. In the DCA, both the XGBoost and logistic regression models provided the greatest net benefit across clinically relevant probability thresholds (0.2-0.6). This ML-based model enables early, individualized prediction of standing ability at discharge among patients with moderate-to-severe TBIs. The inclusion of modifiable variables enhances its clinical utility for rehabilitation planning in the intensive care unit and supports data-driven quality improvement initiatives. External validation is recommended to assess its generalizability.