Abstracts

Rationale: Epilepsy, defined as the risk of recurrent seizures, can negatively affect global development in children. Approximately 30% of newly diagnosed epileptics become refractory to medication based treatment. The quality of life for children can be significantly affected by epilepsy. Due to the severe impact of epilepsy on these children, these refractory epilepsy patients are evaluated for epilepsy surgery workup to collect and analyze key clinical information. This includes seizure semiology (the study of signs that occur during seizures) alongside electrophysiological study, which provides high spatial and temporal resolution monitoring of neural activity. Certain behaviors, such as arm extension or head deviation are associated with certain types of seizures that originate from specific motor regions of the brain. To study epilepsy semiology, trained epileptologists must visually inspect video collected from the Epilepsy Monitoring Unit (EMU) and identify seizure periods. This information is then analyzed alongside recordings of the brain using invasive techniques such as electrocoricography (ECoG) and stereoelectroencephalography (sEEG). This general approach is time consuming and focuses on visually obvious semiologies, leaving nuanced signals relating to seizures to be further studied. Methods: We utilized commercially available, minimally invasive computer vision sensors to augment the collection of normal motor behavior and epilepsy semiology recorded during routine epilepsy surgery workup. The quantified data accumulated across hours upon days from these sensors was analyzed in conjunction with synchronized ECoG or sEEG information to help elucidate neural functioning in specific regions and networks of the brain. We applied semi-automated computer vision algorithms to extract detailed accounts of patient motor behavior from RGB-D videos through dense optical flow and pose estimation. We then applied machine learning techniques from the field of brain-machine interfaces to interpret neural responses to more complex movement behaviors. A similar process was also applied to seizure periods in order to compare both movement and neural responses with respect to non-seizure periods. This approach enables the study of behavioral and neurophysiological correlates that may be predictive of clinically relevant brain regions. Results: We analyzed neural correlates to unconstrained movement behaviors from patients (N=3) in the pediatric EMU. Using the methods mentioned above, we quantified and identified periods of movement activity, some of which were volitional movements whereas others were seizure episodes. We compared each patient’s neural response and movement kinematics between volitional and seizure periods to identify distinguishing characteristics elucidating seizure semiology. Conclusions: This work presents a novel approach to augmenting patient behavioral monitoring during their stay in the EMU using minimally-intrusive computer vision sensors. Unlike traditional EMU based neurophysiology studies that are conducted over 10s of minutes, we record throughout the patient’s EMU stay, generating hours upon days of neural data synchronized to patient video. From these datasets, we can semi-automatically identify and quantify unconstrained behaviors and seizure semiologies that cannot be emulated by structured tasks, thus enabling a big data approach to neural analysis. Funding: Support for this project is funded in part by the Hellman Fellowship, the Institute of Engineering in Medicine Graduate Student Fellowship, and the Clinical and Translational Research Institute at UC San Diego.