Choice of training data for classifier transfer in error related potentials based on signal characteristics
In Proceedings of the 7th International IEEE EMBS Conference on Neural Engineering, (NER-2015), 22.4.-24.4.2015, Montpellier, IEEE, pages 102-105, Apr/2015.
In brain computer interfaces, different amounts of training data can be generated during the same recording time depending on the type of e.g., error related potential (ErrP) that is evoked. In our previous study (Kim & Kirchner, 2013), we obtained more training data containing observation ErrPs compared to interaction ErrPs within the same recording time under similar scenario conditions. Thus we trained a classifier on observation ErrPs to detect interaction ErrPs. This led to the reduction of calibration time. In this previous study we assumed that features extracted from a window, in which both types of ErrPs show a similar shape of averaged activity (0.16– 0.6 s after error events), are optimal for classifier transfer. In this study we test this assumption on a larger group of subjects. Further, we evaluate an extended training window that covers a late negativity at 0.6–0.8 s, which has a stronger Amplitude in case of observation ErrPs. Such an extension of the Training window allows to improve the classification performance in case that observation ErrPs are used to train and test a classifier (no transfer case). However, in this study we will show that for the transfer case this long window [0.16–0.8 s] is outperformed by the short window [0.16–0.6 s], which contains only the part of both types of ErrPs with similar shape. The results indicate that the signal characteristics can guide the choice of Training data for classifier transfer between different types of ErrPs.