Double Stage Transfer Learning for Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signal acquisition for brain-computer interfaces (BCIs) is challenging due to non-stationarity and lengthy calibration.
• Existing transfer learning (TL) methods for EEG often yield suboptimal results by extracting only partial features.

Purpose of the Study:

• To propose a novel double-stage transfer learning (DSTL) algorithm for effective knowledge transfer in BCIs.
• To address limitations of current TL algorithms in EEG feature extraction and domain adaptation.

Main Methods:

• Developed DSTL algorithm applying TL to both preprocessing and feature extraction stages of BCIs.
• Utilized Euclidean alignment (EA) for inter-subject EEG trial alignment.
• Employed covariance matrix reweighting and Common Spatial Patterns (CSP) for spatial feature extraction.
• Applied Transfer Component Analysis (TCA) to minimize domain discrepancies.

Main Results:

• DSTL demonstrated superior classification accuracy on two public EEG datasets.
• Achieved 84.64% and 77.16% accuracy in multi-source to single-target (MTS) paradigms.
• Obtained 73.38% and 68.58% accuracy in single-source to single-target (STS) paradigms.
• Outperformed existing state-of-the-art methods in both transfer paradigms.

Conclusions:

• The proposed DSTL algorithm effectively reduces domain differences in EEG data.
• DSTL offers a promising new approach for EEG data classification without requiring extensive training datasets.
• This method enhances the efficiency and applicability of BCIs.