TabTransformer Prediction of Neurological Deterioration in Mild Traumatic Brain Injury with Computed Tomography

Background: Older patients with traumatic brain injury (TBI) may deteriorate neurologically even when their initial Glasgow Coma Scale (GCS) score is 13–15. This study evaluated whether TabTransformer could predict neurological deterioration within 24 h using clinical findings, coagulation test results, and quantitative computed tomography (CT) measurements obtained at admission. Methods: We retrospectively studied 350 patients who received conservative treatment at a single center. Neurological deterioration was defined as a decrease of at least 2 GCS points within 24 h. A TabTransformer with a multilayer perceptron (MLP) classification head was evaluated against Extreme Gradient Boosting, random forest, logistic regression, and an MLP-only ablation model. Evaluation included the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), sensitivity, specificity, Brier score, and calibration curves. We used Shapley Additive exPlanations (SHAP) to examine model behavior and divided predicted probabilities into tertiles for exploratory risk stratification. Results: On the test set, TabTransformer achieved an AUROC of 0.825 (95% CI, 0.713–0.927), with the event rate reflecting the corrected outcome prevalence. These values were better than those of the comparison models. D-dimer, acute subdural hematoma thickness, fibrinogen, and maximal cerebral contusion diameter contributed most strongly to the predictions. The observed deterioration rate increased across the low-, intermediate-, and high-risk probability groups. Conclusion: In this cohort, the TabTransformer predictions were well calibrated and discriminated between patients who did and did not deteriorate. SHAP values helped explain individual predictions. Validation in prospective cohorts from other centers is required before the model is considered for clinical use.