Sensorimotor rhythms (SMR) have been the neuronal phenomena of choice in non-invasive EEG-based endogenous brain computer interfaces (BCIs) for more than two decades and SMR-based BCIs have achieved the highest degree of freedom control so far. Nevertheless, they are subject to long periods of training prior to attaining a satisfactory level of control requiring users to learn to modulate their rhythms. The goal of this work is to analyse this problem, discuss the causes of the slow rise in performance and provide recommendations on alternative solutions to quicken control attainment.

The study has been conducted by both theoretical and empirical analysis. A theoretical model has been developed that explains the principle operation of SMR-based BCIs focusing on major performance contributors respectively the user, periodic feature selection and the translation model thus contrasting user adaptation and machine learning. Five able-bodied subjects (age:  ) participated in six sessions of online computer cursor control experiments over three weeks to evaluate control attainment performances and gather data for statistical analysis (∼1152 trials per subject). Correlation ( ) between user control features and target position over sessions was assessed as an estimate of neural adaptation and the predictive power of the translation algorithm (10 × 10 fold cross-validation) was calculated over sessions as an estimate of machine adaptation. Auxiliary performance metrics were evaluated.

Features-target correlation increased over sessions, while at the same time the predictive accuracy ( ) of the translation model remained averagely steady and very low ( ) demonstrating continuous user adaptation and low model predictive accuracy. Periodic feature selection was theoretically discussed to be very instrumental and its relevance was empirically illustrated.

The study concludes that the slow control attainment in SMR-based BCIs is due to its reliance on user training (neural adaptation) which is adaptive but too slow in the context of SMR modulations and due to the weak decoding of the neuronal phenomenon utilised by the user. As a recommendation, the optimality of the feature selection algorithm could be looked at to guarantee the use of the most relevant features. However and most importantly the predictive power of the translation model should be significantly improved in order to quicken control attainment as thereafter the control attainment effort could be shifted from neural adaptation to machine learning.