Stroke is the the third most common cause of death and the main cause of disabilities acquired in adults in high-income countries. While recovering from a stroke early mobilisation is an important factor in recovery as well as continuous maintenance of physical conditions for the rest of the patients life. To overcome the disability, several rehabilitation techniques have been proposed, among these is a relatively novel technique: brain computer interface (BCI) this tech- nique either translates brain features into actions or translates features to activate a device. BCI systems have been designed to detect movement related cortical potentials (MRCP) in real time, and other systems have detected not only the onset of movements, but also detected the intended movement speed and force. None of these systems however has investigated how fast BCI control can be learned, they have also not investigated the robustness of the BCI systems over time using MRCP as a control signal.
The aim of this project was to investigate changes in performance over time of a BCI systems designed to detect MRCPs and the intended speed of the movement. It was believed that MRCP could be used as a control signal from the first session, and that the performance of the system would not change significantly during a longitudinal experimental design due to the simplicity of the movements performed.
7 healthy volunteers were used to test what the effect of training of a BCI system over time was and if the MRCP of healthy subjects for non complex dorsiflexions changed over time. The BCI was also tested for one session on 6 patients suffering a stroke affecting their motor cortex to see how the BCI system performed detecting differences in speed based on MRCPs for stroke patients. This was done to test if the BCI could be used for rehabilitation from the first session. A BCI designed to detect MRCPs and classify them based on movement speed, either fast move- ments 0.5 seconds or slow movements 3 seconds was tested during two sessions for four weeks and a for week break followed by a control session. Three different training scenarios were tested to investigate if using data from more than one session would increase the performance of the BCI. The first test focused on the effects of only using previously recorded data, no training of the subject that session prior to the test. The second test only used training data recorded during that session for a more current detector and classifier, but with a smaller amount of data. The third test used all available data and trained the subject in the different types of movements prior to testing, this was believed to have the highest accuracies, but also the most time consuming.
The mean detection, classification, and system performance for the three different test were 79 ± 1 %, 56 ± 1 %, and 45 ± 1 % respectively. A single session was performed with 6 patients to test how well the BCI system worked for patients. The mean detection, classification, and system performance for the three different test were 88 ± 12 %, 57 ± 6 %, and 50 ± 9 % respec- tively.
These findings suggest it is possible to use a BCI with very little training and increase perfor- mance by using all available data. The findings also show no increase or decrease in performance of the system over time, the control session showed that it was not necessary to train the subject after a four week break if enough prior training data was available. The session with the patients showed detection accuracies that can most likely be used help patients rehabilitate by combining the online system with e.g. functional electrical stiulation.