Low-Dimensional Subject Representation-Based Transfer Learning in EEG Decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Advances in passive brain-computer interfaces (BCIs) using electroencephalogram (EEG) show promise for real-world neuromonitoring.
• Human variability in EEG signals presents a significant challenge for practical EEG-based BCI applications.
• Current transfer-learning techniques often require supervised calibration with task-relevant data, which is impractical for real-life scenarios.

Purpose of the Study:

• To present a novel transfer-learning framework for EEG decoding that overcomes limitations of supervised calibration.
• To develop a method for EEG decoding that utilizes low-dimensional subject representations derived from pre-trial EEG data.
• To enhance the practicability and usability of EEG-based BCIs in real-world applications.

Main Methods:

• Applied tensor decomposition to pre-trial EEG data to extract subject, spatial, and spectral characteristics.
• Developed a framework to assess these characteristics and obtain low-dimensional subject representations.
• Identified subjects with similar brain dynamics for effective knowledge transfer.

Main Results:

• The proposed low-dimensional subject representation-based transfer learning (LDSR-TL) framework demonstrated superior prediction accuracy compared to random selection and Riemannian manifold approaches.
• The LDSR-TL framework achieved higher accuracy in cognitive-state tracking.
• The framework requires significantly less training data for new users, enabling rapid, unsupervised calibration.

Conclusions:

• The LDSR-TL framework effectively addresses human variability in EEG data for BCI applications.
• This approach significantly improves the practicability and usability of EEG-based BCIs for real-world neuromonitoring.
• The method facilitates leveraging existing data and minimal new data for accurate BCI development.