EEG correlates and methods for learning in brain-computer interaction.

Motor Imagery (MI)-based Brain-Computer Interface (BCI) has emerged as a promising approach to provide an alternative means of communication, control and rehabilitation for people with severe motor impairments. However, the efficiency and efficacy of BCI systems remain to date rather limited, preventing their out-of-lab implementation. This thesis offers a few stepping stones towards more user-oriented BCI, shifting the focus to subject learning, neuroplasticity monitoring and the co-adaptation between the human and the ML BCI decoder. First, I seek to identify the electroencephalography (EEG) correlates of learning to drive a racing car, an example of complex motor skills. Additionally, I explore the role of anodal transcranial Direct Current Stimulation (tDCS) in enhancing race-driving training. My work determines that theta EEG rhythms and alpha-band effective functional connectivity between frontocentral and occipital cortical areas are salient neuromarkers of the acquisition of racing skills. I also discern a possible tDCS effect in accelerating the pace of learning. My thesis presents a novel feature selection method which combines the conventional data-driven approach with BCI expert knowledge through Fuzzy Logic. I show that my algorithm achieves statistically significant improvement in terms of classification accuracy, feature stability and class bias. The proposed method can promote subject learning during BCI training by keeping the selected features within a “learnable”, physiologically relevant manifold. One of the main motivations behind co-adaptative BCI has been the avoidance of boring and laborious open-loop calibration sessions, imposed at the beginning of user training to collect data for ML BCI model training. For BCI-based rehabilitation, these issues become pressing, demotivating for the patients and hard to fit logistically into a strict clinical schedule. Towards alleviating this issue, this thesis identifies different methods for calibration-free BCI-based rehabilitation. My results indicate that calibration-less BCI-based rehabilitation algorithms are possible without compromising performance. The proposed methods thus lift a major barrier currently obstructing the translation of BCI-based therapies.