Abstracts

Rationale: The automatic detection of seizures with a high degree of sensitivity and specificity is a prerequisite for the development of a responsive closed-loop system for managing seizures in real time. Accurate automated detection can also minimize the need to manually label seizures in large datasets. The development of automated seizure prediction models can be facilitated by a highly reliable detector if the classifications made by the detector are used to train a predictor. Limitations on the size and the low number of seizures recorded per patient in publically available EEG datasets results in classifiers that are accurate internally but do not generalize well. In this work, we explored lightweight models for seizure detection using a highly interpretable feature space in the time and frequency domains on a 1000+ seizure labeled EEG dataset created internally. We then evaluated the performance of this model on a very large seizure-rich dataset obtained from the Wilcox Lab at the University of Utah. Methods: Young wild-type male Sprague-Dawley rats (n = 16) were implanted with surface electrodes, EMG pads, and intrahippocampal depth electrodes bilaterally. Following recovery, epilepsy was induced with microinjections of kainic acid administered via a cannula into the hippocampus. EEG and video was recorded continuously in a rodent epilepsy monitoring unit (EMU). The recordings collected from this experimental setup constituted the Cash lab dataset. This was labeled by an expert and 1012 seizures were recorded from 16 rats. Spike-wave absence seizures were not included. An additional dataset was obtained from the Wilcox Lab at the University of Utah. Rats were recorded continuously via an EcoG channel with video-EEG and seizures were manually labeled by experts. This larger, more varied dataset contained 2883 seizures from 96 rats with multifocal epilepsy also induced by kainic acid injections. EEG data was computed for features in the time and frequency domains for single-channel and multi-channel features (Cash lab dataset) and single-channel features only (Wilcox dataset). We developed pooled and individual generalized linear models (GLMs) with a ridge penalty for detection using the Cash lab dataset and trained a pooled GLM with a ridge penalty for the Wilcox dataset using the same methods. Results: In the Cash lab dataset, analysis of the pooled GLM for single-channel and multi-channel features computed from 3054 ictal windows and 9.96E5 non-ictal windows (trained on 80% of the pooled dataset, tested on the remaining 20%) yielded a test AUROC of 0.995. For the individual classifiers trained on single animals in the Cash dataset, the mean AUROC was 0.962. Extending the pooled GLM to the Wilcox dataset using only single-channel features in the time and frequency domains yielded a test AUROC of .963. Conclusions: We were able to automatically detect seizures with a high degree of sensitivity and specificity at the state of the art with an AUROC of .995 for the pooled GLM on the Cash dataset, using a computationally lightweight method and a highly interpretable feature space. Lower performance on the Wilcox dataset can be attributed to the smaller feature space (25 single-channel features vs. 141 single-channel and multi-channel features). This lends support for the importance of multi-channel features in seizure detection. The low computational complexity of this model makes it a viable candidate for clinical settings. Funding: NIH, Harvard College Research Program, Bertarelli Fellowship, Harvard Medical Scientist Training Program, Paul & Daisy Soros Fellowship