Predicting subject performance level from EEG signal complexity when engaged in BCI paradigm

The ability to monitor or even to predict the performance level of a subject when engaged in a cognitive task can be useful in various real-life scenarios. In this article we focus on a popular EEG-based Brain Computer Interface (BCI) paradigm and report on the complexity of the EEG signals in relation to the subject's performance level. We estimate signal complexity with a multivariate, multiscale version of Sample Entropy (MMSE) to account for multiple temporal scales as well as within and cross-channel dependencies. Furthermore, we apply Multivariate Empirical Mode Decomposition (MEMD) to render the temporal scales data driven instead of predefined. Our pilot study shows that the multivariate entropy of EEG signals changes during the course of the experiment and that it can be used for predicting the subject's performance level (accuracy).