Abstracts

Rationale: Existing automated seizure detection algorithms have been shown to have sensitivities of between 43-63% and specificities between 56-90%. While this represents an improvement over the previous generation of algorithms, more improvements must be made before automated seizure detection can be employed in clinical practice. The algorithms particularly suffer from false alarms when applied to neonatal EEG, due to the high degree of nurse handling and rhythmic patting employed to soothe neonates. Computer vision technology that quantifies movement in real time could distinguish seizure and non-seizure motion, dramatically improving both the sensitivity and specificity of automated neonatal seizure detection algorithms. Methods: We utilized video EEG recordings from neonates undergoing epilepsy workup from clinical EEG systems.  Patients were also enrolled in the NEOLEV2 levetiracetam neonatal seizure treatment trial, which allowed for data storage and retrieval.  IRB approval was granted by UCSD, San Diego, CA. The quantified data from these videos was analyzed in conjunction with synchronized EEG annotated information.  We applied computer vision algorithms to extract detailed accounts of patient and surrounding movement behavior though dense optical flow estimation.   First, using a training set of time-smoothed videos, a spatial optimization method was employed for estimation of optical flow.  Then, we obtained the average spectral density of patient care-generated artifact, which was filtered and a threshold level was set to detection of the artifact event.  Variables to control by the software algorithm included the video sampling frequency of the algorithm, sensitivity of the optical flow velocity filter, range of the spectral bandpass filter, as well as type of optical flow method employed.  Finally, using a testing set of videos, when the filtered optical flow was more than the threshold level, the motion events are decided, and we showed that the optical flow represents the detailed artifact event. Results: We analyzed 320x240 pixel 24 hour-length videos and annotated EEG files (N=43) in the neonatal intensive care unit at our hospital. Using the methods mentioned above, we quantified and identified 197 periods of patting activity, of which 45 generated false positive automated seizure detection events. A binary patting detection algorithm was trained with subset of 362 event videos.  This supervised detection algorithm was applied to a testing subset, which resulted in 37% reduction in false positive automated seizure detection events caused by patient care patting, while maintaining all true positive events. Conclusions: This work presents a novel approach to improving automated seizure detection algorithms used during standard-of-care neonatal video EEG monitoring. When coupled with a neonatal automated seizure detector, our artifact detection mechanism can improve the ability of the seizure detector algorithm to distinguish between artifact and true seizure activity. Funding: This project was partially funded by the UCSD Young Investigator Grant.