Self-Supervised Learning with Adaptive Frequency-Time Attention Transformer for Seizure Prediction and...

Area of Science:

• Deep Learning
• Neuroscience
• Biomedical Signal Processing

Background:

• Improving electroencephalogram (EEG) feature extraction is vital for accurate deep learning-based epilepsy prediction and classification.
• Traditional supervised methods require extensive labeled data, hindering large-scale training.
• Existing self-supervised methods struggle with EEG noise and signal degradation, impacting performance.

Purpose of the Study:

• To develop a self-supervised learning Transformer network with Adaptive Frequency-Time Attention (AFTA) for robust EEG feature representation from unlabeled data.
• To enhance feature extraction robustness against noise in EEG signals.
• To improve the performance of downstream tasks such as seizure prediction and classification.

Main Methods:

• Proposed a self-supervised learning Transformer network with an integrated Adaptive Frequency-Time Attention (AFTA) mechanism.
• Utilized a masking-and-reconstruction framework for pretraining the Transformer network on unlabeled EEG data.
• Incorporated an Adaptive Frequency Filtering Module (AFFM) within AFTA for adaptive frequency domain filtering, combined with temporal attention.

Main Results:

• Achieved state-of-the-art performance across TUSZ, TUAB, and TUEV EEG datasets.
• Obtained highest AUROC (0.891), balanced accuracy (0.8002), weighted F1-score (0.8038), and Cohen's kappa (0.6089).
• Demonstrated robustness and generalization capabilities in seizure detection and classification.

Conclusions:

• The proposed AFTA-enhanced Transformer network effectively learns robust EEG features for epilepsy analysis.
• The method overcomes limitations of traditional and existing self-supervised approaches by mitigating noise impact.
• The findings highlight the potential for improved automated seizure prediction and classification using self-supervised learning.