loading page

EEG-Based Seizure Detection Using Feed- Forward and LSTM Neural Networks Based on a Neonates Dataset
  • Amr Zeedan ,
  • Khaled Al-Fakhroo ,
  • Abdulaziz Barakeh
Amr Zeedan
Author Profile
Khaled Al-Fakhroo
Qatar University

Corresponding Author:[email protected]

Author Profile
Abdulaziz Barakeh
Author Profile

Abstract

Electroencephalography (EEG) signals are used for the diagnosis of neurological disorders and detection of seizures. Early detection of seizures not only can save the patient’s life but also can improve the quality of life. The traditional methods for seizure detection are time-consuming. To combat these problems and further improve the accuracy of seizure detection, the techniques of deep machine learning have been applied to EEG signals. This work develops an EEG-based seizure detection program using deep machine learning to detect seizures using EEG signals. The model is developed using a data set of EEG recordings from 79 human subjects. To this end, a literature review has been conducted to review the latest research related to deep learning algorithms applied to EEG signals for seizure detection. The EEG data has been pre-processed, filtered, and segmented before being fed to the machine learning model. Three models were developed, two models are based on feed-forward neural networks and the third model is based on long-short term memory (LSTM) network. For the feed-forward neural network with 10 hidden neurons, the obtained accuracy was too low (66.1%). For the feed-forward neural network with 100 hidden neurons, the accuracy was slightly improved to 74.3%. The LSTM model showed an accuracy of 87.7%. The model correctly classified 96.4% of normal patients as having no seizures and correctly classified 71.6% of seizure patients as having seizures. The limitations on increasing the accuracy of the models are discussed and possible solutions are suggested. If channel-specific seizure annotations are provided by medical experts, the accuracy of the model is expected to increase significantly.