DCNet: A Self-Supervised EEG Classification Framework for Improving Cognitive Computing-Enabled Smart Healthcare.

Area of Science:

• Cognitive computing and neuroscience
• Machine learning applications in healthcare

Background:

• Electroencephalography (EEG) is vital for cognitive computing models.
• Supervised EEG classification offers high accuracy but requires extensive manual annotation and struggles with generalization.
• Self-supervised models offer an alternative but often fall short of supervised accuracy, facing challenges in temporal dependency capture, loss function design, and data imbalance.

Purpose of the Study:

• To introduce the DreamCatcher Network (DCNet), a novel self-supervised framework for robust EEG classification.
• To address key challenges in self-supervised EEG analysis, including temporal dependencies, loss function adaptation, and data imbalance.
• To improve the accuracy and generalization of EEG classification models for cognitive computing applications.

Main Methods:

• Developed a two-stage training strategy for DCNet, starting with contrastive learning for representation extraction and followed by supervised transfer learning.
• Employed time-series contrastive learning to capture comprehensive temporal correlations within EEG data.
• Introduced SelfDreamCatcherLoss, a novel loss function for evaluating representation similarity, and integrated two data augmentation techniques to mitigate class imbalance.

Main Results:

• Demonstrated the superiority of DCNet over existing state-of-the-art models in EEG classification tasks.
• Achieved high accuracy on both the Sleep-EDF and HAR datasets, validating the framework's effectiveness.
• Showcased DCNet's ability to extract robust representations and handle data imbalance effectively.

Conclusions:

• DCNet presents a significant advancement in self-supervised EEG classification, outperforming current methods.
• The proposed framework shows great potential for revolutionizing sleep disorder detection and advancing cognitive computing in healthcare.
• DCNet offers a promising direction for developing more accurate and generalizable EEG analysis tools.