Identification of Epileptic Seizures using Recurrent Neural Networks and Time Series Transformer

In recent years epileptic seizures have become a major impact on large number of people worldwide. Over 60 million people have been affected globally due to this neurological condition. These seizures are caused by excessive discharge of electrical signals in the brain. To identify the excess discharge of signals, a device called electroencephalogram (EEG) is used. This device captures the signals by placing multiple sensors in scalp of the brain. However, the signals from the brain vary in time with different individuals thereby requiring more data. Also, differences are found in the signals between various stages of seizures. To overcome these issues and to detect the seizure automatically from EEG signal a novel Ensemble approach is proposed using Recurrent Neural Networks (RNN) and Time Series Transformer (TST). Raw EEG signals are pre-processed to remove undesirable noises from the obtained signals for further processing. The TST model was utilized to extract characteristics from EEG recordings. Furthermore, a recurrent neural network incorporating a long input layer, fully connected layer, activation layer was constructed in parallel to train the features, classify the EEG data, and identify the seizures. The effectiveness of the network was further examined and compared using various activation channels. The findings demonstrated that the suggested network could achieve a reasonably accurate prediction using a mere 18 channels. Experiments were carried out on the CHBMIT and Bonn datasets. Proposed method achieved an accuracy of 94.31% with an F-Score of 91.93%. Furthermore, classification of seizures has been improved compared to state of art methods.