Abstracts

Rationale: Post-traumatic epilepsy (PTE) is a life-altering consequence of traumatic brain injury (TBI). Currently, there are no methods for prediction or prevention of PTE. In recent years, investigators have turned towards machine learning models to identify biomarkers of epileptogenesis that can accurately predict the risk of developing PTE. Accurate and sensitive biomarkers can then be used to identify at-risk populations for clinical trials of antiepileptogenic treatments and therapies. This study aims to build off prior work that employed machine learning models using features derived from diffusion weighted imaging to predict the risk of PTE; instead we use structural imaging features extracted from T1-weighted imaging in order to evaluate the comparative predictive power of different structural features. In this study, cortical thickness was chosen as the candidate structural imaging feature for machine learning because cortical thinning has been found to be associated with both TBI outcome and epilepsy in literature.

Methods: Structural MRI images were obtained from 79 subjects enrolled in the Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx). Moderate to severe TBI patients age 6-100 were eligible for EpiBioS4Rx if they presented with frontal and/or temporal hemorrhagic contusion and Glasgow Coma Scale score between 3-13 without continuous sedation at the time of enrollment. Cortical thickness of each patient was extracted using FreeSurfer. A Random Forest (RF) model was created to identify differences in cortical thickness between seizure positive (at least one seizure after seven days of injury) (n = 65) or seizure negative (no seizures) (n = 14). Upsampling was performed for the seizure negative group to reduce bias due the class imbalance. Features that were particularly important for distinguishing the two classes (seizure positive or seizure negative) were recorded.

Results: With over 100 rounds of cross-validation, the RF model employed in this study was able to distinguish between seizure positive and seizure negative patients at a mean accuracy of 86%, sensitivity of 94%, and specificity of 74%.

Conclusions: Our results suggest that this RF model is able to successfully classify alterations in cortical thickness that are associated with seizure activity following TBI. Changes in cortical thickness were most prominent in the left and right rectus gyrus and right temporal sulcus. These findings are consistent with the literature that associates these regions of interest (ROIs) with epileptic activity. Our results provide further support that a multimodal machine learning approach that utilizes both structural T1 imaging and diffusion weighted imaging measures may be even more effective at predicting PTE.

Funding: Please list any funding that was received in support of this abstract.: This work is supported by the National Institute of Neurological Disorders and Stroke (NINDS) at the National Institutes of Health (NIH), award numbers U54NS100064 and R01NS111744.