Poster Z11 - Towards automatic seizure detection within zebrafish larvae using single channel electroencephalography

Introduction: Presently, epilepsy treatment primarily consists of patients taking anticonvulsant medications to suppress seizures. While many patients experience a reduced number of seizures, approximately a third of patients have intractable, or drug-resistant, epilepsy and do not benefit from anticonvulsant medications. As such, researchers continue to pursue new drugs and therapies capable of treating epilepsy with resistance to extant medications. One animal model utilized to quicken the rate of development for new treatments is zebrafish (Danio rerio) due to their beneficial properties as a model organism. However, for epilepsy research in both human and animal models, such as zebrafish models, labeling electrophysiological data is a lengthy procedure requiring trained technicians to comb through hours of recordings. Therefore, an investigation was performed into creating automatic seizure detection algorithms for electroencephalography (EEG) data recorded from zebrafish epilepsy models.

Materials and

Methods: Single-channel EEG recordings from seizure-induced zebrafish epilepsy models were performed and used to develop two seizure detection algorithms that detect seizure, neurotypical, and artifact-corrupted data from signal features within 4-second windows of the recordings. These seizure detection algorithms were based on k-nearest neighbor (KNN) and artificial neural network (ANN) multi-class classifiers. A 10-fold cross-validation procedure with 10% of the data reserved for validation was used to evaluate the performance of the detection algorithms.

Results, Conclusions, and Discussions: The KNN detection algorithm demonstrated a classification accuracy of 80%, slightly higher than the 77.31% accuracy of the ANN classifier. These initial findings serve as a proof-of-concept for the potential of automatic seizure detection using only a single-channel EEG, paving the way for future research. Both algorithms showed robust performance when labeling neurotypical data but struggled somewhat when distinguishing between seizure and artifact-laden data. This suggests a potential strategy of splitting the labeling task into two separate binary decisions: the first one labeling data as neurotypical or non-neurotypical, and the second one determining if non-neurotypical data are from seizures or artifacts. We also suspect that incorporating information from adjacent windows could significantly enhance the accuracy of the algorithms. To our knowledge, this is only the second attempt to automatically label seizure single-channel EEG data from zebrafish, and the first to do so without removing neurotypical data from recordings manually beforehand. These preliminary results are encouraging and will serve as a starting point for continued research.

Acknowledgements (Optional):