Early risk stratification for in-hospital mortality and 6-month unfavourable prognosis after injury in traumatic brain injury: machine learning models for clinical decision support

Abstract Purpose To develop and externally validate machine-learning models using routinely available early admission data for in-hospital death and 6-month unfavourable prognosis after injury in traumatic brain injury (TBI). Methods This retrospective multicentre prognostic study included patients aged ≥ 15 years admitted within 12 h after TBI at four Chinese hospitals from November 2020 to October 2025. Two hospitals were used for development and two for independent external testing. Predictors included age, neurological assessments, laboratory measurements, and head computed tomography (CT) findings. Laboratory measurements were standardised using Value/upper limit of normal (ULN). Models were selected internally and evaluated using discrimination, classification, calibration, and decision curve analysis. Results The in-hospital death analysis included 2260 patients (523 deaths, 23.1%); the 6-month unfavourable prognosis after injury analysis included 1512 discharge survivors with available follow-up (254 unfavourable outcomes, 16.8%). For in-hospital death, CatBoost using features selected by Lasso-Logistic regression achieved external area under the receiver operating characteristic curve (AUROC) 0.903 (95% confidence interval [CI], 0.883–0.922), area under the precision–recall curve (AUPRC) 0.800 (95% CI, 0.760–0.844), sensitivity 82.0%, positive predictive value (PPV) 71.1%, and F1 score 76.2%. For 6-month unfavourable prognosis after injury, logistic regression using features selected by XGBoost-based recursive feature elimination with cross-validation (RFECV-XGBoost) achieved external AUROC 0.906 (95% CI, 0.873–0.936), AUPRC 0.727 (95% CI, 0.642–0.823), sensitivity 81.1%, PPV 72.5%, and F1 score 76.6%. Conclusion Both models maintained consistent cross-centre predictive performance and may support early risk stratification and treatment planning after TBI. Prospective multicentre evaluation is still needed before implementation.