Abstract:
Background: Severe traumatic brain injury (sTBI) in adults remains a leading cause of mortality and disability worldwide. Early identification of reliable predictors of outcome is crucial for risk stratification and ICU management. Disturbances of hemostasis and metabolic factors such as body mass index (BMI) have been proposed as potential prog- nostic markers, but evidence remains limited. Methods: We conducted a retrospective, multicenter study including 307 adult patients with sTBI (Glasgow Coma Scale ≤ 8) admit-ted to three tertiary intensive care units in Ukraine between September 2023 and July 2024.All patients underwent surgical evacuation of hematomas and decompressive craniotomy. Laboratory parameters (APTT, INR, fibrinogen, platelets, D-dimer) were collected within 12 h of admission. BMI was calculated from measured height and weight. Predictive modeling was performed using L1-regularized logistic regression and Random Forest algorithms. Class imbalance was addressed with SMOTE. Model performance was as-sessed by AUC, accuracy, calibration, and feature importance. Results: The 28-day all-causemortality was 32.9%. Compared with survivors, non-survivors had significantly lower GCS scores and higher INR, D-dimer, and APTT values. Very high VIF values indicated severe multicollinearity between predictors. Classical logistic regression was not estimable due to perfect separation; therefore, regularized logistic regression and Random Forest were applied. Random Forest demonstrated higher performance (AUC 0.95, accuracy ≈ 90%) than logistic regression (AUC 0.77, accuracy 70.1%), although results must be interpreted cautiously given the small sample size and potential overfitting. Feature impor- tance analysis identified increased BMI, prolonged APTT, and elevated D-dimer as leading predictors of mortality. Sensitivity analysis excluding BMI still yielded strong performance (AUC 0.91), confirming the prognostic value of coagulation markers and GCS. Conclusions: Mortality in adult sTBI patients was strongly associated with impaired hemostasis, obesity, and low neurological status at admission. Machine learning-based modeling demonstrated promising predictive accuracy but is exploratory in nature. Findings should be interpreted with caution due to retrospective design, severe multicollinearity, potential overfitting, and absence of external validation. Larger, prospective, multicenter studies are needed to confirm these results and improve early risk stratification in severe TBI.
