Abstract
INTRODUCTION: Traumatic Brain Injury (TBI) is one of the leading causes of mortality and severe disability worldwide. This study aimed to develop and optimize machine learning (ML) algorithms to predict abnormal brain computed tomography (CT) scans in patients with mild TBI. METHODS: In this retrospective analyses, the outcome was dichotomized into normal or abnormal CT scans, and univariate analyses were employed for feature selection. Then SMOTE was applied to address class imbalance. The dataset was split 80:20 for training/testing, and multiple ML algorithms were evaluated using accuracy, F1-score, and area under the receiver operating characteristic curve (AUC-ROC). SHAP analysis was used to interpret feature contributions. RESULTS: The data included 424 patients with an average age of 40.3 ± 19.1 years (76.65% male). Abnormal brain CT scan findings were more common in older males, patients with lower Glasgow Coma Scale (GCS) scores, suspected fractures, hematomas, and visible injuries above the clavicle. Among the ML models, XGBoost performed best (AUC 0.9611, accuracy 0.8937), followed by Random Forest, while Naive Bayes showed high recall but poor specificity. SHAP analysis highlighted that lower GCS scores, decreased SpO2 levels, and tachypnea were strong predictors of abnormal brain CT findings. CONCLUSION: XGBoost and Random Forest achieved high predictive accuracy, sensitivity, and specificity. GCS, SpO2, and respiratory rate were key predictors. These models may reduce unnecessary CT scans and optimize resource use. Further multicenter validation is needed to confirm their clinical utility.