Intersession adaptation of the EEG-based detector of self-paced walking intention in stroke patients

Brain-computer interfaces (BCIs) have been used in patients with motor impairments as a rehabilitation tool, allowing the control of prosthetic devices with their brain signals. Typically, before each rehabilitation session a calibration phase is recorded to account for session-specific signal changes. Calibration is often an inconvenient process due to its length and patients' fatigue-proneness. This paper focuses on improving the performance of an EEG-based detector of walking intention for intersession transfer. Nine stroke subjects executed a self-paced walking task during three sessions, with one week between sessions. We performed an intersession adaptation by using 80% of one session's data and an additional 20% of a next session for training, and then we tested the detection model on the remaining part of the next session. In practice, this would constitute a longer initial calibration (40 minutes) and a shorter recalibration in subsequent sessions (10 minutes). After training set adaption we attain an average increase in performance of 13.5% over non-adaptive training. Furthermore, we used an approximation of Kullback-Leibler (KL) divergence to quantify the difference between training and testing sets for the non-adaptive and adaptive transfer. As a potential explanation for the improvement of intersession performance, we found a significant decrease in KL-divergence in the case of adaptive transfer.