Efficient recognition of event-related potentials in high-density MEG recordings

In brain-computer interfacing (BCI), the recognition of task-specific event-related potentials such as P300 responses is an established approach to regaining communication in severely paralyzed people. However, a reliable detection of single trial potentials is challenging, because they are strongly affected by noise. Furthermore, potentials with their subcomponents are often distributed over several channels. With high density sensor arrays, a hypothesis-driven selection of channels, as often performed in BCIs based on electroencephalography (EEG), is challenging. We present a new data-driven approach that constructs spatio-temporal filters, considerably reducing the number of channels, reducing noise, and simultaneously determining the underlying brain dynamics. The extracted signals can be easily used to recognize the event sequence on which users focus their attention, without applying multivariate classification. We evaluated the approach using high density magnetoencephalography (MEG) data, recorded during a BCI experiment based on P300 responses. Compared to the subject's performance achieved with the initial decoding approach, the recognition rate increased significantly from 74.1% (std: 14.8%) to 95.1% (std: 4.9%) correct detections, which implies an information transfer rate improvement from 6.9 bit/min to 13.1 bit/min on average over 17 subjects.