Latent variable method for automatic adaptation to background states in motor imagery BCI

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interface (BCI) systems are sensitive to user background state variations.
• Lack of user background state information during training limits BCI performance.
• Unsupervised methods are needed to address background state variability in BCIs.

Purpose of the Study:

• To develop and evaluate an unsupervised method for incorporating background states into BCI systems.
• To improve the classification accuracy of target states in BCIs.
• To enable recognition of background states without requiring labeled data.

Main Methods:

• Proposed a probabilistic latent variable model with a discrete latent variable.
• Utilized the expectation-maximization algorithm for parameter estimation.
• Applied the method to asynchronous motor imagery data from 12 subjects with open/closed eyes as background states.

Main Results:

• The latent variable method improved target state classification in 7 out of 12 subjects compared to a baseline.
• The method successfully recognized background states in 6 out of 12 subjects.
• Demonstrated effectiveness in handling unlabeled background state data.

Conclusions:

• The latent variable method offers a robust approach to enhance BCI classification accuracy by integrating unsupervised background state information.
• This method improves BCI performance by considering background states during training and prediction.
• Provides a significant advancement for developing more reliable and adaptive BCI systems.