Bayesian Signal Matching for Transfer Learning in ERP-Based Brain Computer Interface

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-Computer Interface (BCI) spellers use electroencephalogram (EEG) signals to aid communication for individuals with disabilities.
• P300 Event-Related Potentials (ERPs) are key EEG signals for BCI spellers, but current calibration methods are time-consuming and can reduce accuracy.
• Existing calibration relies on individual participant data, leading to lengthy training and potential user fatigue, which biases P300 estimation.

Purpose of the Study:

• To develop a novel Bayesian signal matching (BSM) framework for efficient Brain-Computer Interface (BCI) calibration.
• To reduce the lengthy training time associated with traditional BCI calibration methods.
• To improve the prediction accuracy of BCI spellers by leveraging data from source participants.

Main Methods:

• Proposed a Bayesian signal matching (BSM) framework utilizing a Bayesian hierarchical mixture model.
• Specified the joint distribution of stimulus-specific EEG signals among source participants.
• Implemented an inference strategy where similar participants share model parameters, while dissimilar ones retain unique parameters.

Main Results:

• Demonstrated the advantages of the BSM framework through simulations.
• Successfully applied BSM to real-world data from participants with neuro-degenerative diseases.
• Showcased the framework's ability to generalize to other base classifiers with parametric forms.

Conclusions:

• The proposed Bayesian signal matching (BSM) framework offers a significant improvement over existing Brain-Computer Interface (BCI) calibration strategies.
• BSM effectively reduces calibration time and enhances prediction accuracy by utilizing data from source participants.
• This approach holds promise for improving communication accessibility for individuals with severe motor impairments, particularly those with neuro-degenerative diseases.