Improving Deep Learning for Seizure Detection using GAN with Cramer Distance and a Temporal-Spatial-Frequency Loss Function

Abstract

The signals of EEG are analyzed in the identification of seizure and diagnosis of epilepsy. The visual examination process of EEG data by skilled physician is huge time-utilization and the judgemental process is complicated, which may vary or show inconsistency among the physician. Hence, an automatic process in diagnosis and detection was initiated by the Deep Learning (DL) approaches. Time Aware Convolutional Neural Network with Recurrent Neural Network (TA-CNN-RNN) was one among them. Deep neural networks trained on large labels performed well on many supervised learning tasks. Creating such massive databases takes time, resources, and effort. In many circumstances, such resources are unavailable, restricting DL adoption and use. In this manuscript, Generative Adversarial Networks with the Cramer distance (CGAN) is proposed to generate an accurate data for each lable. A spatiotemporal error factor is introduced to differentiate actual and genetrated data. The discriminator is learned to differentiate the created data from the actual ones, while the generator is learned to create counterfeit data, which are not estimated as false by the discriminator. The classical GANs have a complex learning because of the nonlinear and non-stationary features of EEG data which is solved by Carmer Distance in the proposed method. Finally, the sample generated by CGAN is given as input for the Time Aware Convolutional Neural Network with Recurrent Neural Network (TA-CNN-RNN) classifier to investigate experimental seizure Prediction outcome of the proposed CGAN. From the investigational outcomes, the proposed CGAN- TA-CNN-RNN model attained classification accuracy of 94.6%, 94.8% and 95.2% on CHB-MIT-EEG, Bonn-iEEG and VIRGO-EEG than other existing EEG classification schemes and also provides great potentials in real-time applications.