Adaptive Classification Improves Control Performance in ERP-Based BCIs

This contribution investigates the effects of applying an unsupervised adaptation mechanism to linear classifiers for Brain-Computer Interfaces (BCI). Specifically, we track changes in the first two moments of the unlabeled data distribution. Changes are adaptively compensated by recalculating the classifier based on short, consecutive data segments. The approach is validated on three auditory oddball data sets containing a total of N=37 subjects, of which 6 were used for model selection and the remaining 31 for validation. We find a significant performance increase (up to 14%) for the adaptive scheme compared to a fixed classifier. The increase is largest for subjects with low performance.