Abstracts

Rationale: Ictal semiology is used to diagnose epilepsy, classify seizure type, and monitor the efficacy of anti-epileptic drugs. Using a porcine model of post-traumatic epilepsy (PTE), we aim to identify, classify, and quantify a variety of epileptic semiologies with the goal of identifying behavioral markers of PTE, and perhaps behavioral markers of epileptogenesis. We used manual analysis to inform automated analysis on video of control swine and swine with PTE

Methods: Yucatan pigs received bilateral cortical impact (n = 10) or sham surgery (n = 3) at an average of 5 months of age. Video was recorded for an average of 11.5 months for 24 hours per day as animal housing space allowed. New video camera arrays were tested in an additional 2 injured pigs. Pigs were classified as having PTE based on the presence of recurrent convulsive seizures. Manual analysis was performed on about 5,000 hours of video to build a library of behaviors that might be greater in pigs with PTE. Using DeepLabCut (DLC), nine distinct body parts: tail, each leg, forehead, snout, and both ears were labeled from sample frames via an aerial video camera. These labels were then utilized by a training process for supervised machine learning. 10,000 iterations on 120 frames were utilized to determine body part coordinates. 

Results: Epileptogenesis in this species was prolonged occurring over several months resulting in approximately 70,000 hours of video among all subjects. Manual analysis of a subset of videos resulted in a library with 15 distinct behaviors including freezing, head bobbing or wet dog head shakes, and forelimb clonus that were identified as potentially more prevalent in pigs with PTE vs. pigs without PTE. Tonic limb extensions and myoclonic/hypnotic jerks were not distinguishable from normal sleep behavior. Manual analysis was time consuming and fraught with low inter-rater reliability, and therefore, machine learning was pursued. A total of 12 seconds of a roughly 46-minute video was labeled with 9 body parts and DLC successfully extrapolated body part locations on the remaining unlabeled video frames. Preliminary results indicate that DLC can track body part movements efficiently in a swine model of PTE. Next, video files containing potential automatisms identified via manual analysis will be labeled with DLC, classified with Simple Behavioral Analysis, and quantified in pigs with PTE or without PTE. Additionally, we will quantify machine-identified automatisms and patterns and compare in those with or without PTE.  

Conclusions: In a swine model of PTE, DLC provides an advantageous solution for automatically estimating body part coordinates. Next, we aim to use body part coordinates from DLC as an input for classification program to identify behavioral biomarkers of PTE in swine, and perhaps, predict the development of PTE.

Funding: DOD, CURE W81XWH-15-2-0069, NIH NICHD K01HD083759, NIH R01 HD099397, The Claflin Distinguished Scholar Award