Adversarial Temporal Modeling for Seizure Classification Using WGAN-GP-LSTM Networks

Detecting an epileptic seizure with the help of electroencephalogram (EEG) signals is difficult due to the absence of classified data or complexity. The conventional GAN-based data augmentation methods suffer from instability and inadequate learning of dependencies, which limits their efficiency despite increasing the diversity of data. An LSTM method is proposed for binary seizure classification with the WGAN-GP-LSTM scheme to overcome the previous issues. The generator and the discriminator of the system proposed in this paper use LSTM units, which differentiate it from other previously proposed systems. The system can produce acceptable EEG signals. This also helps in the correlations being learned over time for seizure classification. Through this system, EEG datasets that contain signals with and without supplements are publicly available. Results indicate that using synthetic data augmentation will enhance the WGAN-GP-LSTM model’s performance. Consequently, this model outperforms WGAN-GP and other prior works on the overall score, which are 0.96, 0.875, 0.82, and 0.923, respectively, for overall accuracy, sensitivity, specificity, and F1 score. WGAN-GP, with an FID score of 3.07 is poorer than WGAN-GP with a score of 1.42. The WGAN-GP-LSTM model displaces adversarial sequence generation and classification, helping to overcome challenges of data scarcity and the temporal dependence of EEG signals for seizure detection. A different approach to computer -assistance in neurodiagnostics has been introduced.