Reliable subject-adapted recognition of EEG error potentials using limited calibration data

For the development of efficient Brain Computer Interfaces (BCIs), recognizing when the system reacts erroneously to a user´s input is a much desired functionality. In this paper, we investigate a system for the recognition of error potentials from single-trial Electroencephalography (EEG). Our focus here is the development of a system using only limited calibration data from the test subject, while exploiting available training data from other subjects. In an evaluation with 20 sessions, we show that we can achieve an average F-score of up to 0.86 for a system using ICA-based artifact correction and training data filtering which only requires few minutes of additional calibration data.