Skip to Main Content
Brain-computer interface (BCI) as a rehabilitation tool has been used in restoring motor functions in patients with moderate to sever stroke impairments. To achieve the best possible outcome in such an application, it is highly desirable to have a stable and accurate operation of BCI. However, since electroencephalogram (EEG) signals considerably vary between sessions of even the same user, typically a long calibration session is recorded at the beginning of each session. This process is time-consuming and inconvenient for stroke patients who undergo long-term BCI sessions with repeating same mental tasks. This paper investigates the possibility of omitting the intra-session calibration for BCI-based stroke rehabilitation when large data recorded from the same user are available. For this purpose, a large dataset of EEG signals from 11 stroke patients performing 12 BCI-based stroke rehabilitation sessions over one month is used. Our offline results suggest that after recording a number of stroke rehabilitation sessions, the patient does not require calibration any more. The experimental results show that combining 11 sessions, which each session comprises minimum 60 trials per class, yields a model that averagely outperforms the standard calibration model trained by the data recorded directly before the test session.