Classifier Transferability in the Detection of Error Related Potentials from Observation to Interaction

A challenge in brain computer interface (BCI) applications is the reduction of time required for the acquisition of training data, which is needed for a user specific calibration of a BCI. This paper proposes an application oriented approach to minimize the calibration time by transferring a classifier between different types of error related potentials (ErrPs). A classifier trained to detect a certain brain pattern is used to later detect a brain pattern which is expected to be similar to the first one. In the here presented approach, two different tasks (interaction task/observation task) are performed within one scenario which is developed to generate two types of ErrPs: interaction ErrPs and observation ErrPs. Since almost twice as much training data can be generated while performing the observation task compared to the interaction task within the same calibration time, we use the classifier trained on the data containing observation ErrPs to evaluate it's performance on the data containing the interaction ErrPs. Presented results support our approach. We show that a single trial detection of interaction ErrPs using the classifier trained on observation ErrPs is possible and results on average in a high detection performance of 0.77 balanced accuracy [(TPR+TNR)/2], i.e., an average of recognition rate of correct and erroneous trials of 77%. Without such classifier transfer, the classification performance of observation and interaction ErrPs is on average slightly higher (0.80 and 0.81 balanced accuracy, respectively). Our results suggest that classifier transfer is feasible and reduces calibration time. This is a relevant result from the perspective of applying an ErrP-based brain-computer interface in a realistic scenario in robotics.