On the feasibility of compressing intracranial EEG with Generative Adversarial Networks

The recording of intracranial electroencephalogram (iEEG) signals is essential in neuroscience but produces large amounts of data due to their high temporal resolution and dimensionality. Efficient compression methods are therefore required, yet the variability of iEEG across subjects and channels makes this task particularly challenging. This thesis builds on a previous study that explored Generative Adversarial Networks (GANs) for iEEG compression which introduces an encoder within the architecture to produce latent representations of the signals. Training follows a two-phase procedure: joint training of encoder and GAN, and optimization-quantization of the latent space while refining the GAN. We introduced a clustering strategy as a preprocessing step. The algorithm divides channels from a single patient into groups according to their dynamical similarity. Then, different compressors capture the specific characteristics of each group. As a preliminary step, we tested several GAN architectures on their ability to generate synthetic iEEGs from noise. Among these, a Temporal Convolutional Network GAN (TCN-GAN) and a Transformer-based GAN (TGAN) achieved satisfactory results in generation, and we decided to test them for the compression task. The results show that clustering produces consistent groups, but their heterogeneous sizes affect compression performance and limit scalability. The generation phase proved useful for model selection but did not ensure reconstruction capabilities. We did not identify a final model capable of successfully compressing and reconstructing iEEG signals. This confirms the complexity of the task and the difficulties of training GANs, where optimization, hyperparameter tuning and stability required extensive effort with results often limited to specific configurations.