To identify computed tomography (CT) based radiomic biomarkers of post-traumatic epilepsy (PTE) in traumatic brain injury (TBI) patients and develop machine learning (ML) models for PTE risk prediction.

Radiomics, the extraction of quantitative features from medical images, is a rapidly evolving field in neuroimaging research. By leveraging radiomic features extracted from brain CT scans and robust machine learning models, this study aims to identify novel neuroimaging biomarkers that could predict the incidence of PTE following a TBI.

Patients with severe TBI treated at a level 1 trauma center (Houston, TX) and ≥3 months of clinical follow-up were included. Features included CT-based radiomics, demographics, social determinants, and clinical outcomes. The study cohort was randomly divided a priori 80:20 into training and evaluation. Random Forest (RF) and Logistic Regression (LR) models from Scikit-Learn and XGBClassifier (XGB) and XGBRFClassifier (XGBRF) from XGBoost were tuned with nested cross-validation and evaluated. Analyses were repeated to minimize bias, and feature importance was assessed to identify biomarkers. 

Eighty-two patients (58 male, 24 female; median age 43.5) were analyzed, with 48% experiencing PTE. Of the models tested, LR had the best performance (AUC 0.722 ± 0.082) across all runs. Three radiomic biomarkers that recurred across model iterations in greatest to least stability were frontal lobe zone variance, limbic lobe coarseness, and frontal lobe large-area low-grain-level emphasis. Collectively, these features suggest that regional heterogeneity and large-scale structural texture patterns within the frontal and limbic lobes are key imaging correlates of clinical outcomes in TBI in the context of PTE.

This study identifies radiomic features of potential clinical significance and shows that machine learning models hold promise for generating meaningful prognostic insights in TBI. LR achieved moderately strong performance and revealed that textual heterogeneity and large-area gray-level patterns in the frontal and limbic lobes were robust radiomic biomarkers.