A comparative study on adaptive subject-independent classification models for zero-calibration error-potential decoding

Today, a substantial part of human interaction is the engagement with artificial technological and information systems. Error-related potentials (ErrPs) provide an elegant method to improve such human-machine interaction by detecting incorrect system behaviour from the electroencephalography (EEG) signal of a human operator or user in real time. In this paper, we focus on adaptive subject-independent classification models particularly suitable for the task of ErrP decoding. As such, they provide a promising method to overcome the need of individualized decoding models, which require a time consuming calibration phase. In a comparative study we evaluate the performance of a decoding model solely trained on prior data and the effectiveness of adapting this model to a new subject. Our results show that such a generalized model can decode ErrPs with an acceptable accuracy of $(72.73\pm 5.27){\%}$ and that supervised adaptation can significantly improve the accuracy of the generalized model. Unsupervised adaptation did only prove useful for some subjects with high initial model accuracy and requires more sophisticated methods to be practical for a broader range of subjects. Consequently, our work contributes to the development of calibration-free ErrP decoding, which can potentially be used to improve human-robot interaction.