Post-Adaptation Effects in a Motor Imagery Brain-Computer Interface Online Coadaptive Paradigm

Abstract

Online coadaptive training has been successfully employed to enable people to control motor imagery (MI)-based brain-computer interfaces (BCIs), allowing to completely skip the lengthy and demotivating open-loop calibration stage traditionally applied before closed-loop control. However, practical reasons may often dictate to eventually switch off decoder adaptation and proceed with BCI control under a fixed BCI model, a situation that remains rather unexplored. This work studies the existence and magnitude of potential post-adaptation effects on system performance, subject learning and brain signal modulation stability in a state-of-the-art, coadaptive training regime inspired by a game-like design. The results extracted in a cohort of 20 able-bodied individuals reveal that ceasing classifier adaptation after three runs (approx. 30 min) of a single-session training protocol had no significant impact on any of the examined BCI control and learning aspects in the remaining two runs (about 20 min) with a fixed classifier. Fifteen individuals achieved accuracies that are better than chance level and allowed them to successfully execute the given task. These findings alleviate a major concern regarding the applicability of coadaptive MI BCI training, thus helping to further establish this training approach and allow full exploitation of its benefits.

Transfer learning promotes acquisition of individual BCI skills

“…Yet, they only partially showed that online decoder adaptation supports BCI skill acquisition as just selected subjects were discussed in terms of their final EEG patterns. Nevertheless, other studies found that online decoder adaptation did not support subject learning even if BCI performance improved ( 44 , 45 ). Intermittent decoder recalibration using data from previous sessions also promoted subject learning ( 13 , 15 , 29 ), although they required longer number of training sessions than the approaches reported in this work.…”

“…Supervised approaches that use labeled samples when recalibrating the BCI decoder are popular ( 16 , 18 , 41 , 43 , 45–47 ), as ML theory predicts that they would lead to better model fitting and performance—a result verified in a cross-over BCI study comparing supervised vs. fully unsupervised approaches for online decoder adaptation ( 44 ).…”

EEG changes during passive movements improve the motor imagery feature extraction in BCIs-based sensory feedback calibration

“…Various research have demonstrated the positive impact of adding MI EEG data from periods in which users received sensory feedback, to calibrate the BCI [19]. Is the BCI's output completely correlated to the user's intention by using this strategy?…”

Workshops of the eighth international brain–computer interface meeting: BCIs: the next frontier

“…These metrics should also inform earlier stage BCI development before end-user evaluation [139,140]. Further factors should be considered when designing the BCI paradigm, for instance, the design of tasks, feedback, instructions, and signal processing [86,[141][142][143]. Performance may improve via engaging task design (e.g.…”