Bayesian Adaptive Stimulus Optimization in P300 Brain Computer Interfaces

Stimulus-driven brain-computer interfaces (BCIs) present stimuli to users in order to evoke neural responses that are used as control signals for external devices. The stimulus presentation schedules for these BCIs are typically designed pseudo-randomly. However, a data-driven approach enables stimulus selection to be informed by recently recorded neural activity from the BCI user. We seek to optimize BCI stimulus presentation schedules in order to enhance the information contained in neural responses that are elicited from presented stimuli and thereby increase BCI communication rates.

We have developed an adaptive decision framework for P300-based BCIs that uses an information-based criterion to dynamically generate stimuli in near-real time to maximize BCI communication efficiency. We impose constraints in the search space during stimulus selection to account for data processing delays and to mitigate the negative impact of psychophysical factors, such as refractory effects and distraction effects, during real-time BCI use.

Our adaptive stimulus selection paradigm achieves improved communication rates (non-significant decrease in spelling accuracy, significantly improved spelling speed, non-significant increase in bit rate) relative to the conventional checkerboard stimulus presentation paradigm in an online P300 speller study.

For computational simplicity in implementation, the optimal stimulus can generally be approximated by finding the stimulus that has a combined probability mass closest to 0.5.

We present a practical implementation of an adaptive stimulus selection paradigm that is driven by online data and improves BCI communication speed.