The BCI, however, is also sensitive to changes in the user's underlying mental states, such as fatigue, frustration and attention level. "At our hospital, we have completed three paediatric case studies, and have several others on their way. From our experience, the effect of changing mental states can be quite impactful," said Tom Chau, vice-president of research at Holland Bloorview Kids Rehabilitation Hospital (a fully affiliated academic health sciences centre of the University of Toronto).

"For example, in one of our patient studies, we had a severely disabled nonverbal individual who became very excited with the feedback provided, realizing that he was exerting control over the computer," Chau explained. "The subsequent brain activation that accompanied what we suspected was a combination of novelty and reward processing threw off the BCI. In our able-bodied studies we have seen overt frustration and self-reported fatigue in particular play a non-trivial role in diminishing BCI performance."

To mitigate this problem, Chau and his PhD student Andrew Myrden investigated a psychologically adaptive EEG-BCI that adapts to short-term mood changes. The system combines a passive BCI, which predicts the user's current mental states, and an active BCI that differentiates between two mental tasks.

Myrden and Chau examined two adaptive BCI approaches. In the first, they used predicted mental states to assess the likelihood of task misclassification by a non-adaptive BCI. In the second, the predicted states were used to directly adapt the BCI and increase task classification accuracy (J. Neural Eng. 13 066022).

Training and testing

The study included 11 able-bodied individuals who used a task-based EEG-BCI to play a simple maze navigation game. They first completed training sessions comprising thirty trials of four active mental tasks and an unconstrained rest task. BCIs were trained to differentiate active tasks from the rest task using a linear discriminant analysis classifier. One active task was selected for each individual, and the BCI that best differentiated that task from the rest task was used during testing.

Participants then undertook three testing sessions, during which they performed maze navigation or the rest task whilst self-reporting their perceived levels of fatigue, frustration and attention. During each session, a wireless EEG headset recorded electrical activity from 15 locations on the cerebral cortex. EEG signals for each task were used to generate a set of 450 features, based on the power within 30 frequency ranges for each of the 15 electrodes. Features chosen from this set were then employed for mental state prediction, BCI reliability prediction and BCI adaptation.

The researchers compared each subject's self-reported mental state levels with those predicted using least squares regression. The average Pearson correlation coefficient between reported and predicted values was 0.46 for attention, 0.50 for fatigue and 0.56 for frustration, demonstrating that the mental state prediction algorithm was moderately accurate for the population.

They then used these mental state predictions to estimate BCI reliability. Applying a threshold of 0.4 to differentiate reliable from unreliable tests revealed an 8% difference in the BCI's accuracy for trials flagged as reliable and those deemed unreliable - implying that mental state-based prediction of BCI reliability exceeded chance levels.

The adaptive BCI

Next, Myrden and Chau designed an adaptive BCI that retrained a new classifier for each testing sample using only training samples with similar predicted mental states to those of the testing sample. They compared the classification accuracy of the adaptive BCI to that attained by two non-adaptive BCIs, one that randomly sampled the training set to match the size of the adaptive training set, and one that used the entire training set.

In the first case, for most participants, the adaptive BCI outperformed the non-adaptive BCI for smaller training set sizes, approaching the performance of the non-adaptive BCI when the entire training set was used. In the second comparison, the adaptive BCI generated improved performance for five participants.

Finally, they used algorithms to identify the ideal combination of mental states and training set size for each participant. On average, each mental state was selected approximately 80% of the time. When this combination was used to adapt the BCI classifier, five participants exhibited significant improvements in classification accuracies, while no participants exhibited significantly worse performance. Overall, the adaptive BCI exhibited an accuracy of 73.2% while the non-adaptive BCI exhibited an accuracy of 72.6%.

Myrden and Chau concluded that mental state predictions can be used both to estimate BCI reliability and directly improve classification accuracy. These findings represent a first step towards the design of BCIs that can adapt to short-term changes in psychological state.

"As a next step, we'd like to better understand when adaptation would be beneficial and arrive at specific criteria to guide the invocation of adaptation – i.e., by how much does the user's psychological state need to change before we trigger BCI adaptation?" Chau told medicalphysicsweb.

Related stories

• Microstimulation could restore lost touch
• Implant restores hand motion after paralysis
• Patient's thoughts control robotic arm
• FES helps stroke patients move again