Semiautomatic physiologically-driven feature selection improves the usability of a brain computer interface system in post-stroke motor rehabilitation

In an electroencephalographic (EEG)-based BCI-assisted Motor Imagery (MI) training the reinforcement of a specific EEG pattern elicited by correct MI requires that expert neurophysiologists with knowledge of BCI technology identify the optimal control features for each single patient. This procedure is highly dependent on the operator and is currently restricted to researchers with experience in the BCI field and specific neurophysiological knowledge. To increase the usability of BCI technology and thus foster its use in clinical routine, we developed a semiautomatic method to select control features by combining both physiological and statistical approaches. The aim of the study is to compare classification performances obtained using BCI control features selected by expert professional user (manual procedure) and those obtained by semiautomatic method (guided procedure). The application of the guided procedure on real data sets showed performances comparable to those obtained with manual procedure. This suggests that it is feasible to successfully support the professional endusers, such as therapist/clinicians who are not necessarily expert in BCI field, in the EEG feature selection yet according to evidence-based rehabilitation principles