Towards Calibration-Less BCI-Based Rehabilitation

Brain-computer interfaces (BCIs) are increasingly deployed in stroke rehabilitation. Most BCIs rely on batch, supervised parameter estimation which requires large, labelled brain data associated with the motor tasks to be performed during the therapy. Consequently, BCI-based rehabilitation regimes include an initial calibration session aimed at collecting the necessary data to train the BCI model. This calibration process is time-consuming and tedious, especially considering the strict logistical constraints in a clinical setting. This paper investigates the possibility of calibration-free BCI regimes rendering recalibration sessions entirely redundant. We compare the decoding performance of three different approaches to calibration-less BCI, which exploit either the notion of adaptation or the classical results of event-related synchronization/desynchronization, to that of a conventional, calibration-based classification method, on a large dataset of 26 (sub-)acute stroke patients performing 15 therapeutic sessions. Our results show that calibration-less BCI for stroke treatments is not only possible, thus lifting a major practical barrier hindering the translation of this technology to clinics, but may also be superior to the standard, calibration-based methodology in terms of classification accuracy.