SCH: New Statistical Learning Methods for Brain-Computer Interfaces

Brain-computer interfaces (BCIs) are an emerging communication and computer access option for people with severe physical impairments, such as those who are in a "locked-in" state due to an acquired or congenital disability. One of the most successful non-invasive BCIs for communication is the P300-BCI design, named after the brain activity that is called the P300 event-related potential (ERP).

This model works by presenting stimuli (flashing groups of keys) to an on-screen keyboard. The individual's electroencephalogram (EEG) after each stimulus is then classified according to whether it contains the ERPs produced only for the stimulus the user has selected as their target. This ERP-based BCI design can be calibrated for an individual in a single session.

However, the BCI still takes time for user calibration and the selection speed is slow. The calibration process has been especially challenging for people without other communication methods and for children, who have limited attention spans. This project will create new statistical methods that 1) reduce the time required to calibrate the BCI for an individual user, 2) reduce the calibration effort for individual users by leveraging prior knowledge from other BCI users, and 3) improve the selection speed of the BCI through dynamic adjustments to the patterns of stimuli.

This contribution is significant because the proposed methods will substantially improve the classification process and communication speed. The research outcome of this project will also provide new insights for a better understanding of brain functions and neurobiology of thinking, and provide valuable information for the future design of the BCI system.

The team will involve undergraduate and graduate students in the project, educate them on state-of-the-art research in BCI and statistical machine learning, organize summer training workshops, and develop free software.

This project will develop a series of statistical methods and study their theoretical properties for analyzing brain signals from BCI systems and making statistical inferences about brain activity. The project will focus on three unique but related problems. First, the project will establish a dynamic statistical learning framework for analyzing BCI brain signals, including the split-and-merge Gaussian process for classification, a new logistic stick-breaking process for synchronization, and the novel information-guided autoencoder for extracting the latent factors in the signals.

Second, the team will address BCI data integration problems, such as combining EEG from multiple users by utilizing subgroup identifications to capture the heterogeneity in the brain activity across the population, and integrating useful prior knowledge, such as brain functional connectivity networks, into statistical inferences on brain responses. Finally, the project will develop a reinforcement learning method that dynamically adjusts the presentation of groups of stimuli by the BCI for optimal identification of the target stimulus, based on the development of a Markov decision process and the Q-learning method via deep neural networks.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.