Automatic detection of seizures within zebrafish larvae epilepsy models using single-channel electroencephalography

Abstract

Researchers continue to pursue new drugs capable of treating intractable, or drug- resistant, epilepsy as a large number of patients do not see a reduction in the number of seizures from current treatments. To quicken the pace of drug research, zebrafish (Danio rerio) have been utilized as a model organism for testing anticonvulsant drugs before clinical trials. However, the lengthy task of labeling electroencephalography (EEG) data slows the pace of this line of research and limits its full potential. This study investigated the creation of automatic seizure detection algorithms for electroencephalogram data recorded from seizure-induced zebrafish to detect seizure, artifact, and neurotypical events. Four unique seizure detection algorithms were proposed and implemented based on k-nearest neighbor (KNN), support vector machine (SVM), and artificial neural network (ANN) classifiers. These four techniques were tested using the same input features and their results compared. The best-performing algorithm was identified as the Single Stage KNN with an 83.8% accuracy, followed by the Single Stage ANN with an 80.3% accuracy. The results indicate that a single-stage, multiclass classification architecture may be beneficial to automatically labeling epilepsy data, thus enhancing efficiency. Furthermore, the results for the algorithms which separate the multiclass classification into a series of binary classifications suggest that there are advantages in research and clinical settings to implement a detection algorithm that can delineate neurotypical and non-neurotypical data to assist with manual labeling.