Large-Scale Assessment of a Fully Automatic Co-Adaptive Motor Imagery-Based Brain Computer Interface

Abstract

In the last years Brain Computer Interface (BCI) technology has benefited from the development of sophisticated machine leaning methods that let the user operate the BCI after a few trials of calibration. One remarkable example is the recent development of co-adaptive techniques that proved to extend the use of BCIs also to people not able to achieve successful control with the standard BCI procedure. Especially for BCIs based on the modulation of the Sensorimotor Rhythm (SMR) these improvements are essential, since a not negligible percentage of users is unable to operate SMR-BCIs efficiently. In this study we evaluated for the first time a fully automatic co-adaptive BCI system on a large scale. A pool of 168 participants naive to BCIs operated the co-adaptive SMR-BCI in one single session. Different psychological interventions were performed prior the BCI session in order to investigate how motor coordination training and relaxation could influence BCI performance. A neurophysiological indicator based on the Power Spectral Density (PSD) was extracted by the recording of few minutes of resting state brain activity and tested as predictor of BCI performances. Results show that high accuracies in operating the BCI could be reached by the majority of the participants before the end of the session. BCI performances could be significantly predicted by the neurophysiological indicator, consolidating the validity of the model previously developed. Anyway, we still found about 22% of users with performance significantly lower than the threshold of efficient BCI control at the end of the session. Being the inter-subject variability still the major problem of BCI technology, we pointed out crucial issues for those who did not achieve sufficient control. Finally, we propose valid developments to move a step forward to the applicability of the promising co-adaptive methods.

Intra- and Inter-subject Variability in EEG-Based Sensorimotor Brain Computer Interface: A Review

“…Around 15-30% users are inherently unable to produce taskspecific signature robust enough to control a BCI (Blankertz et al, 2009;Vidaurre and Blankertz, 2010). The underlying causes of this BCI illiteracy are not well-understood; however, diverse psychological and neurophysiological predictors appear to be associated with BCI performance (Blankertz et al, 2009;Vidaurre and Blankertz, 2010;Jensen et al, 2011;Hammer et al, 2012;Ahn and Jun, 2015;Jeunet et al, 2015;Reichert et al, 2015;Zhang et al, 2015;Acqualagna et al, 2016;Vasilyev et al, 2017;Sannelli et al, 2019).…”

“…Cognitive and neurological factors including functions and anatomy along with emotional and mental processes give rise to intra-and inter-subject variability affecting the performance of SMR-based BCI (Wens et al, 2014;Reichert et al, 2015;Zhang et al, 2015;Acqualagna et al, 2016;Betzel and Bassett, 2017;Vasilyev et al, 2017;Seghier and Price, 2018;Betzel et al, 2019;Smith et al, 2019). Time-variant cognitive factors such as fatigue, memory load, attention and reaction time modulate instantaneous brain activity, and can cause inconsistent SMRbased BCI performance (Hammer et al, 2012;Ahn and Jun, 2015;Fox et al, 2015;Jeunet et al, 2015;Darvishi et al, 2018;Sannelli et al, 2019).…”

“…RSNs represent large-scale spatiotemporal structures exhibiting intrinsic brain activities that are thought to be functionally relevant (Deco et al, 2011). Studies have shown intra-and inter-subject variability in sensorimotor RSN, which may have implications for BCI performance variability (Jensen et al, 2011;Wens et al, 2014;Reichert et al, 2015;Zhang et al, 2015;Acqualagna et al, 2016;Vasilyev et al, 2017). It has been hypothesized that SMR-based BCI performance predictor is reliable for people who display strong resting EEG amplitudes (Blankertz et al, 2010;Suk et al, 2014;Sannelli et al, 2019).…”

“…Figure 1 schematically illustrates the idea of covariate shift when the training and test data distributions are different. The underlying time-variant and subject-specific brain dynamics depends on associated psychological and neurophysiological factors (Blankertz et al, 2009;Vidaurre and Blankertz, 2010;Jensen et al, 2011;Hammer et al, 2012;Ahn and Jun, 2015;Jeunet et al, 2015;Reichert et al, 2015;Zhang et al, 2015;Acqualagna et al, 2016;Vasilyev et al, 2017;Sannelli et al, 2019) and cause covariate shift in EEG-derived feature distributions (Krusienski et al, 2011;Fazli et al, 2015;Jayaram et al, 2016).…”

Riemannian geometry for EEG-based brain-computer interfaces; a primer and a review

“…The major factors jeopardizing the operation of a BCI decoder in real-world operation are the inter-subject physiological variability and the great variability of real-world environmental conditions (e.g., [56]). Three inter-connected lines of research are trying to overcome these limits, addressing jointly the improvement of BCI usability and robustness:…”