Adaptive Sequence-Based Stimulus Selection in an ERP-Based Brain-Computer Interface by Thompson Sampling in a Multi-Armed Bandit Problem

A Brain-Computer Interface (BCI) is a device that interprets brain activity to help people with disabilities communicate. The P300 ERP-based BCI speller displays a series of events on the screen and searches the elicited electroencephalogram (EEG) data for target P300 event-related potential (ERP) responses among a series of non-target events. The Checkerboard (CB) paradigm is a common stimulus presentation paradigm. Although a few studies have proposed data-driven methods for stimulus selection, they suffer from intractable decision rules, large computation complexity, or error propagation for participants who perform poorly under the static paradigm. In addition, none of the methods have been applied to the CB paradigm directly. In this work, we propose a sequence-based adaptive stimulus selection method using Thompson Sampling in the multi-bandit problem with multiple actions. During each sequence, the algorithm selects a random subset of stimuli with fixed size, aiming to identify all target stimuli and to improve the spelling speed by reducing the number of unnecessary non-target stimuli. We compute “clean” stimulus-specific rewards from raw classifier scores via the Bayes rule. We perform extensive simulation studies to compare our algorithm to the static CB paradigm. We show the robustness of our algorithm by considering the constraints of practical use. For scenarios where simulated data resemble the real data the most, the spelling efficiency of our algorithm increases by more than 70%, compared to the static CB paradigm.