Seizure Detection Using the Empirical Mode Decomposition and Domain Adaptation

Epilepsy is a group of neurological disorders characterized by recurrent epileptic seizures. By visualizing the EEG (Electroencephalogram) recordings, experts can diagnose the type of seizure and the damaged area of the brain to initiate an antiepileptic drug thereby reducing the risk of future seizures. Some recent methods have been devised to deal with the inherent drawbacks of inter-individual variability; however, some issues still exist as challenges for researchers. In this work, five intrinsic mode functions (IMFs) are extracted, and the domain adaptation (DA) method is used to get common features among different patients. To be more specific, the maximum mean discrepancy-adversarial autoencoders (MMD-AAE) are developed to minimize the inter-domain distance in a high-dimensional space. This analysis is carried out by two datasets, one is the public dataset from CHB-MIT, and one is obtained from Nguyen Tri Phuong general hospital located in Ho Chi Minh City, Viet Nam. Among fifteen quality seizure files with associated notations in Nguyen Tri Phuong general hospital, three recordings containing onset seizures are labelled by the doctors. These data are suitable for public datasets from CHB-MIT, consequently bringing many advantages for testing the model. The accuracy of the model reached 92% for datasets of CHB-MIT and 87% for Nguyen Tri Phuong general hospital datasets.