Abstracts

Rationale: Rodent preclinical models serve a key role in epilepsy research, enabling us to evaluate potential therapeutics and identify disease mechanisms. Determining the frequency, duration, and severity of seizure events – the seizure yield – is an essential component of such work. However, the state of the art requires expensive and labor-intensive methods (i.e., electroencephalography, manual seizure scoring) that limit the pace and scale of experimentation. Thus, there is a need for research tools that reduce the experimental burden associated with seizure analysis. In previous work, we showed that piezoelectric motion sensors on the cage floor show potential for non-invasive seizure detection. Here we expand on this work, utilizing a sensor configuration that is external to the cage, and test the feasibility of seizure detection in a home-cage setting.

Methods: Forty female Wistar rats (6-8 weeks old) underwent acute status epilepticus (SE) for 90 minutes using the lithium-pilocarpine model. At least three months after SE, rats were monitored individually in standard cages using piezoelectric Adapt-a-Base sensors (Signal Solutions, LLC) along with continuous digital video. Week-long piezoelectric signals were processed to derive a simple line length feature time series, and the 40 highest values were identified for each animal. Timestamps of the peaks were stored to a spreadsheet for subsequent review of the video record, and annotated according to observed behavior (i.e., seizure, grooming, etc). The detection precision was then computed as the number of true seizures detected relative to the number of candidates reviewed for each week.

Results: Signals from piezoelectric sensors were found to capture the essential dynamics of seizures, as well as pre- to post-ictal changes in breathing and signatures of other behaviors (freezing, tremor, etc.). The derived detection statistic was successful in detecting seizure events of severity S3-S5 on the Racine scale, albeit with variable precision (23.3 ± 3.8%; mean ± S.E.M.) – possibly a consequence of reviewing a fixed number of candidates per week rather than applying a predetermined threshold based on foreknowledge of the feature distribution for seizure and non-seizure data.

Conclusions: While the true performance of this method cannot be fully characterized without either EEG analyses or exhaustive video review, the method was successful in that it allowed us to identify a large number of seizure events, which will be used as training data for further refinement. Future efforts will focus on exploring additional seizure-related signal features and temporal dynamics, and more sophisticated machine learning approaches in the hope of improving detection sensitivity and specificity. Overall, this approach has potential to significantly reduce the labor associated with identifying spontaneous, tonic-clonic seizures in chronic epilepsy monitoring as compared to exhaustive video review, and could be a tool of value to the research community.

Funding: This work was supported by National Institutes of Health Grant No. NS107148 (to SS) and R01NS079507 (to BB).