Producing synthetic signals from EEG signals with deep learning

In this thesis, synthetic EEG signals were generated using deep learning methods using the Bonn University EEG dataset, and the similarity of these signals with the original signals was evaluated comprehensively. In the study, EEG signals representing different clinical conditions were analyzed using Classic GAN, one of the deep learning algorithms, and the similarities of the synthetic signals to the original signals in terms of amplitude, frequency and time series properties were tested through correlation matrices, RMSE, MAE, MSE and Spectral similarity. It was observed that the synthetic signals showed high similarity to the original signals. It was observed that the similarity rates, especially in groups C and D, were quite high. Due to the long duration of the EEG signals, segmentation process was applied to the data. Each signal was divided into sub-parts of 256 samples. This process increased the efficiency during the training of the model and provided a better understanding of certain frequency bands. The hyperparameters used in the training process were determined as batch size 64, the learning rate 0,000001, epoch 5000 and latent size 256. These settings allowed the generator to learn more complex signals and increased the ability of the discriminator to distinguish fake data. The generator and discriminator losses were monitored during the training process, showing that the model made the synthetic data more realistic and improved its ability to distinguish near-realistic data. Graphical analyses reveal that the quality of the synthetic signals increased as the training process progressed and the system experienced a balanced learning process. The results show that in this study, significant success was achieved in the process of generating synthetic EEG signals with GAN techniques and the generated signals showed high similarity to the original signals. In the future, different architectures and data augmentation techniques can be used to improve model performance. It can contribute to a better understanding of the potential of artificial intelligence-supported systems in medical applications.