Event-driven signal model and active recursive intent estimation for brain-computer interfaces

Abstract

Brain-Computer Interface (BCI) systems can provide a new pathway of communication and control that can be used in both medical and non-medical domains. Electroencephalogram (EEG) signals have been shown to be effective in inferring user intent in BCI applications. However, in many cases, EEG-based communication lacks sufficient accuracy and speed due to three major limitations: 1) inefficient learning process (querying) to gain information for inference, 2) excessive querying to surpass a hard pre-defined threshold, and 3) insufficient classification performance to estimate the user intent. To enhance the performance of BCIs, we have formulated an EEG-based BCI system in a recursive Bayesian state estimation (RBSE) framework. Accordingly, a new sequential active query selection method has been proposed for the non-invasive EEG-based BCI typing systems, to enhance the typing speed and accuracy. The intended symbol estimation requires a selection of queries to be presented to the user. Inspired by the multi-arm bandit framework, we have presented a tractable solution to the subset query selection problem. More specifically, a new action-value function has been introduced for query selection, which uses a linear combination of two objective functions, 1) mutual information between the intended state (user-intended symbol) and the EEG evidence, and 2) momentum which is a function of logarithmic changes of the posterior probability across sequences. We have analytically verified that the proposed policy provides higher probability in picking the intended state for the query subset compared to other commonly used methods. Motivated by the findings for the query selection problem, we have proposed a new stopping criterion based on a history-based objective. We formulated the stopping objective as a linear weighted combination of the posterior and the proposed momentum term. We present an empirical evaluation for both proposed methods, in simulation, and in a human-in-the-loop experiment for a language-model-assisted EEG-based BCI typing system called RSVP Keyboard. Furthermore, we have proposed a method to improve the classification procedure. Typically, there are two major factors which adversely affects the classification performance, 1) inadequate training EEG trials and 2) ignoring the temporal dependency of these trials, especially for rapid series paradigms. We propose a parametric sequence-based model for EEG signals to factor in temporal dependency of trials and to generate synthetic data to augment the size of the training set. We consider all trials that induce brain responses in a rapid-sequence fashion, including a mixture of consecutive target and non-target trials. Using a superposition of a time-invariant process with an auto-regressive (AR) process, EEG signals are treated as a linear combination of a stationary Gaussian process and time-locked impulse responses to the stimuli (input events) onsets. This results in significantly fewer parameters to be estimated, compared to modeling the EEG feature vector for each trial assuming independence (trial-based model).