Adaptive deep feature representation learning for cross-subject EEG decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Collecting large electroencephalogram (EEG) datasets for robust model training is challenging due to time and labor constraints.
• Limited data hinders the generalizability of EEG decoding models, necessitating solutions like domain adaptation.
• Current domain adaptation methods for EEG decoding often struggle with residual domain shift, leading to misclassifications.

Purpose of the Study:

• To propose a novel adaptive deep feature representation (ADFR) framework for enhancing cross-subject EEG classification.
• To improve the generalizability and accuracy of EEG decoding models, particularly in low-data scenarios.
• To develop a method that learns transferable EEG feature representations effectively.

Main Methods:

• The ADFR framework employs maximum mean discrepancy (MMD) regularization to minimize domain distribution discrepancies.
• Instance-based discriminative feature learning (IDFL) regularization is used to enhance the discriminative power of learned features.
• Entropy minimization (EM) regularization is integrated to refine classifier decision boundaries, improving performance through synergistic learning.

Main Results:

• The ADFR framework demonstrated effectiveness on two public motor imagery (MI)-based EEG datasets (BCI Competition III 4a and IV 2a).
• ADFR achieved average accuracy improvements of 3.0% and 2.1% over state-of-the-art methods on the respective datasets.
• The results indicate significant enhancements in cross-subject EEG classification performance.

Conclusions:

• The proposed ADFR algorithm is effective for EEG decoding, showing significant improvements in cross-subject classification.
• The framework demonstrates potential for practical applications in brain-computer interfaces and other EEG-based technologies.
• The study highlights the benefits of combining domain alignment, discriminative feature learning, and entropy minimization for robust EEG decoding.