Generative Adversarial Network for short-term EEG signals compression

Medical signals obtained through EEG (Electroencephalography) are commonly used to diagnose and treat different illnesses such as epilepsy and seizure disorders. These signals are very complex due to their sensitivity to noise and their difficult understanding. Additionally, they tend to have a big size which makes difficult their processing and utilization. The EEG signals consist of multiple channels depending on the number of electrodes that are used to measure brain activity. This means that the size of the signals is a function of the measurement time and the number of channels. Based on these premises, compression is needed to facilitate the transport and storage of the signals. However, compression is a challenging task because there can be information losses in the process. Moreover, it is difficult to detect if the data has been affected after the compression. In recent years, computing power has improved significantly allowing the development of new models used for compression. In particular, the Autoencoders and the Generative Adversarial Networks (GANs) have been used to compress images. Based on this, this work uses a GAN architecture to compress and reconstruct images generated from EEG signals. To prove that the data after the compression is similar to the original one, a RESNET 18 classifier was used with both the original and the reconstructed data to compare if the evaluation metrics are close to each other. The similarity in the performance indicators could suggest that the reconstructed images conserve the main features identified by the classifier. Overall, the results show that the GAN can be effectively used for EEG signals compression. It was observed that the more compressed the images are, the higher the losses after the compression. In particular, if the images are compressed to a size that is more that 8 % of the original size, the classifier behaves similarly with original and reconstructed data. Therefore, the reconstructed data is close to the original one. In contrast, with a compression percentage below 8 % the values of classification accuracy in the reconstructed data are significantly lower than the accuracy with the original images.