查看文章

This thesis presents a self-supervised deep learning approach for learning and extracting representations from long-period electroencephalogram (EEG) input data, in order to be used in downstream tasks such as prediction of dementia, Alzheimer's, general abnormality, or any other long-period and instance-level downstream task. The proposed method employs multi-view contrastive learning and Transformer-based architecture to extract useful representations from raw EEG data in both time and frequency domains. The study investigates the use of unlabeled data augmentations in conjunction with Transformers for the goal of feature representation learning and the combination of different views of time and frequency for effective pre-training tasks. The developed model is evaluated and validated using pre-training and downstream prediction tasks, demonstrating promising results in encoding long-period EEG data, as well as using the resulting representations for condition prediction. This research aims to contribute to the advancement of deep learning techniques in the analysis of EEG data and has potential applications in the early detection and diagnosis of neurological disorders, and opens the door for further research and investigation in this area.