Abstract / Summary
Traumatic brain injury (TBI) is a major cause of disability and death worldwide. In patients with severe TBI, early identification of those less likely to demonstrate neurological improvement during hospitalization may assist postoperative monitoring and risk stratification. This study compared multiple machine learning algorithms for predicting in-hospital Glasgow Coma Scale (GCS) improvement after surgery and explored the relative importance of perioperative clinical variables. We retrospectively analyzed 196 adult patients with severe TBI who underwent surgical treatment between July 2019 and June 2023. A total of 25 perioperative variables were included, covering demographic characteristics, laboratory indices, intraoperative vital-sign summaries, operative variables, and injury-type indicators. Thirteen machine learning algorithms were evaluated, including logistic regression (LR), Naive Bayes, XGBoost, BP neural network, support vector machine (SVM), K-nearest neighbors (KNN), LightGBM, ExtraTrees, CatBoost, gradient boosting decision tree (GBDT), AdaBoost, random forest (RF), and decision tree. Model selection was primarily based on discrimination in internal cross-validation, and calibration of the final model was further assessed in a held-out internal test set. Patients in the improvement group (IM, n = 102) were younger than those in the non-improvement group (NIM, n = 94) (median age, 44 vs. 57 years; p < 0.001). In internal cross-validation, RF achieved the highest area under the receiver operating characteristic curve (AUC = 0.955), indicating the best overall discrimination, whereas XGBoost yielded the highest accuracy (0.885) and F1-score (0.884). In the internal test set, the final RF model showed good discrimination (AUC = 0.935) and acceptable calibration (Brier score = 0.097). Postoperative blood glucose was the most important feature in the final model (21.3%), followed by blood urea nitrogen (10.0%), age (8.8%), minimum systolic blood pressure (6.4%), and platelet count (5.3%). The random forest model showed good internal performance for predicting in-hospital GCS improvement after surgery in patients with severe TBI, with postoperative glucose emerging as the most influential predictor. These findings suggest that perioperative glucose-related factors may contribute to early neurological risk stratification. However, the model remains preliminary and requires external validation in larger multicenter cohorts before clinical implementation.