Introduction

Traumatic brain injuries (TBIs) affect 2.5 million Americans annually, with over 300,000 severe injuries resulting in emergency room visits and hospital admissions.¹ Posttraumatic epilepsy (PTE), or recurrent seizures more than one week after TBI, is a debilitating complication of TBI. Estimated incidence ranges from 2%-50% largely based on severity of injury, among other factors.² The latency period prior to seizure onset allows researchers to investigate biomarkers of epileptogenesis that could predict which patients are most susceptible to developing PTE.

Prior work has investigated whether phenotypic characteristics of TBI lesions on magnetic resonance imaging (MRI) and computed tomography (CT) are related to PTE outcome following TBI. For example, contusions and temporal lobe localization have been associated with increased risk of PTE.³ One recent study of 57 TBI patients also showed how using imaging features in combination with demographic and clinical variables can increase seizure classification performance.⁴ The goal of this work is to establish whether multimodal lesion phenotyping can help classify seizure outcomes in the Epilepsy BioInformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx).

Methods

EpiBioS4Rx enrollment criteria include frontal and/or temporal hemorrhagic contusion and Glasgow Coma Scale (GCS) score between 3-13 without continuous sedation at the time of enrollment. CT and T2-Weighted-Fluid-Attenuated Inversion Recovery (T2-FLAIR) MRI were acquired for 146 EpiBioS4Rx subjects. Manually traced lesion segmentations on T2-FLAIR were produced using ITK-SNAP with two class labels for lesion core and edema (Fig.1).⁵ Lesion segmentations were validated by a board-certified physician with expertise in neuroradiology.

Seizure outcomes were available for 59 patients, with binary labels for seizure-positive (at least one seizure after 7 days post-injury) and seizure-negative (no seizures within the 2-year follow-up window). The seizure-negative group was composed of 35 subjects (29 male, 6 female) with average age=40.46 (SD=21.4), average GCS=8.97 (SD=4.57), and average MRI acquisition at 10.0 (SD=7.48) days post-injury. The seizure-positive group was composed of 24 subjects (19 male, 5 male) with average age=44.21, average GCS=7.54 (SD=3.84), and average MRI acquisition at 13.54 (SD=7.17) days post-injury. For both groups, CT was acquired on the day of hospital admission.

In order to explore the relative importance of different variable categories, 29 variables were placed into 4 groups: a) demographics (age, sex, days post-injury of the MRI acquisition; b) clinical data (GCS at hospital admission); c) MRI-based characteristics extracted from lesion segmentations (voxel- and volume-wise size of lesion core and edema, total lesion size, lesion-to-edema ratio, number of independent lesions); and d) CT-based characteristics describing the types (hemorrhagic, subcortical, non-hemorrhagic, probable brain laceration, deep brain structure, cortical), hemispheres (left or right), and locations of contusions (Frontal, Parietal, Temporal, or Occipital lobe; Internal Capsule; Thalamus/Basal Ganglia, Midbrain, Pons, Medulla, Cerebellum).

Five binary logistic regression models were designed to distinguish seizure outcome: Model 1 used demographic information only; Model 2 used demographic and clinical data; Model 3 used demographic, clinical, and MRI-based lesion data; Model 4 used demographic, clinical, and CT-based contusion characteristics; Model 5 used demographics, clinical data, MRI-based lesion data, and CT-based contusion characteristics.

Results

Model sensitivity, specificity, and area under the curve (AUC) were as follows: Model 1 - sensitivity=85.7%, specificity=33.3%, AUC=0.656; Model 2 - sensitivity=82.8%, specificity=41.7%, AUC=0.701, Model 3 - sensitivity=85.7%, specificity=50%, AUC=0.787, Model 4 - sensitivity=82.9%, specificity=62.5%, AUC=0.808, Model 5 - sensitivity=91.4%, specificity=75%, AUC=0.886 (summary in Fig.2). A confusion matrix for Model 5 is shown in Fig.3, and AUC performance for all models is shown in Fig.4.

Discussion

The highest model sensitivity, specificity, and AUC for seizure classification was found from Model 5, which combined demographic, clinical, MRI, and CT variables. Although sensitivity remained relatively consistent across models, the specificity and AUC notably increased with the addition of imaging features. Increasing specificity is especially critical when considering these results in the context of identifying patients as candidates for antiepileptogenic therapies: a high false positive rate would propose patients with low PTE risk as good candidates for potentially unnecessary interventions.

The high performance of Model 5 also suggests that a multimodal approach to identifying imaging biomarkers of epileptogenesis may yield synergistic results. Frontal and temporal contusions are a widely recognized risk factor for PTE, but the addition of variables describing the full scope of lesion burden seems to add additional context that improves the specificity of seizure outcome classification. This improvement may also be attributed to the inclusion of imaging features acquired at both acute (day of hospital admission) and more chronic (~10-14 days post-injury) time points.

Conclusions

Introducing both CT- and MRI-based characteristics describing TBI phenotypes improved seizure outcome classification. Multimodal phenotyping that includes whole-brain lesion characteristics and nuanced radiological annotations regarding contusions will ideally produce highly sensitive and specific seizure prediction.

Acknowledgements

Our work is supported by NINDS U54 NS100064, and is disseminated on behalf of The Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx) investigators.