Optimizing Deep Learning Models for Real-Time Seizure Detection Based on Electroencephalogram

Accurate and real-time seizure detection is crucial for epilepsy management. Deep learning offers promising solutions, but optimal performance in clinical settings requires efficient and accurate models. This paper focuses on optimizing deep learning models for real-time seizure detection from EEG signals. We evaluate and optimize various architectures on the TUH Seizure dataset, focusing on combining model and architecture in order to leverage strengths, hyperparameter tuning and feature extraction. Our optimized MobileNEp (Mobile Network for Epilepsy) model demonstrates superior efficiency and general seizure detection, reaching an AUC of 0.9363. Optimized Transformers models show better specialization for specific seizure types, achieving AUCs of 0.987 for absence and 0.991 for tonic-clonic seizures. This research highlights the importance of model optimization in developing effective deep learning tools for real-time EEG seizure detection in clinical settings.