Epilepsy is the most prevalent neurological disorder in the world. Although epilepsy has been recognized for centuries, clinical doctors still lack reliable automated tools to diagnose epileptic seizures in electroencephalograms (EEGs). The research community has made significant efforts to develop automated systems for identifying and quantifying epileptic seizures, with many studies reporting excellent accuracy. However, clinicians continue to rely on manual annotations because automated techniques exhibit poor generalization performance when applied to EEG data from new patients. Another challenge in the field is translating the results of preclinical studies conducted on animals to clinical applications in humans.

This work contributes to both challenges. Firstly, we investigate the reasons behind the lack of generalization in automatic models. We find that most existing techniques are evaluated on seizure classification tasks, while clinical doctors primarily encounter detection tasks in their practice. We demonstrate that the performance of automated pipelines differs significantly between the two and identify the key distinction between the tasks: classification presumes a prior separation between seizure and non-seizure EEG signals, whereas detection requires no such prior knowledge. Secondly, we bridge the gap between preclinical and clinical studies by developing novel deep learning architectures. Our best model, trained on EEG data from epileptic mice, demonstrates excellent generalization with an F1-score of 93% when tested on human data.