Dynamic source domain selection: An adaptive EEG transfer learning framework for mitigating negative transfer.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for brain-computer interfaces (BCIs).
• Existing transfer learning (TL) methods in motor imagery BCIs (MI-BCIs) struggle with merging diverse data and inter-subject variability, leading to negative transfer.
• This limits the effectiveness and real-world applicability of current MI-BCI systems.

Purpose of the Study:

• To develop an adaptive EEG dynamic transfer learning framework for MI-BCIs.
• To overcome the limitations of existing TL methods in handling diverse source domains and inter-subject variability.
• To enhance the accuracy and robustness of MI-BCIs through effective cross-domain knowledge transfer.

Main Methods:

• Utilized wavelet transform convolution for time-frequency decomposition of EEG signals.
• Implemented a dynamic adaptive matching strategy using a migration-based attention module to align source and target domain features in a latent space.
• Employed a novel combined loss function to minimize both marginal and class-conditional discrepancies.

Main Results:

• Achieved classification accuracies of 78.78% on BNCI2014001, 82.11% on BNCI2014002, and 78.19% on BNCI2015001.
• Demonstrated robustness to inter-subject variability.
• Outperformed baseline algorithms with accuracy improvements ranging from 0.13% to 27.7%.

Conclusions:

• The proposed adaptive framework effectively addresses the domain shift challenge in MI-BCIs.
• Enables robust cross-domain knowledge transfer, significantly improving the practical application of EEG-based BCIs.
• The approach enhances MI-BCI system performance by bridging distribution disparities between different data domains.