Pretrained Transformers for Seizure Detection

Abstract

Epilepsy is a neurological disorder characterized by seizures that can disrupt a patient’s quality of life. EEG has been used to detect underlying neural activity for diagnosis and treatment. However, standard methods of seizure detection are time-consuming and require manual detection by a trained clinician, with poor inter-clinician agreement. Automated analysis of EEG data offers the potential to improve diagnostic accuracy and reduce manual error. Here, we introduce a transformer-based model pretrained using annotated EEG scalp data that can detect the presence of seizures in a behind-the-ear wearable device setup for the 2023 ICASSP Signal Processing Grand Challenge. Our model has high sensitivity (100%) and a low False Alarm rate (1.78 FA/hr) for seizure detection on a hold-out test set. We also demonstrate how boosting gamma power in preprocessing can improve the performance of ChronoNet, an established EEG abnormality detection model.

Publication
In IEEE International Conference on Acoustics, Speech and Signal Processing
Justin Chen
Justin Chen
Ph.D. Student