Brain-machine interfaces (BMIs) use brain signals to control closed-loop systems in real-time. This comes with substantial challenges, such as having to remove artifacts in order to extract reliable features, especially when using electroencephalography (EEG). Some approaches have been described in the literature to address online artifact correction. However, none are being used as a 'gold-standard' method, and no research has been conducted to analyze and compare their respective effects on statistical data analysis (inference-based decision).
Objective: In this paper, we evaluate methods for artifact correction and describe the necessary adjustments to implement them for online EEG data analysis.
Approach: We investigate the following methods: artifact subspace reconstruction (ASR), fully online and automated artifact removal for brain-computer interfacing (FORCe), online empirical model decomposition (EMD), and online independent component analysis. For assessment, we simulated online data processing using real data from an auditory oddball task. We compared the above methods with classical offline data processing, in their ability (i) to reveal a significant mismatch negativity (MMN) response to auditory stimuli; (ii) to reveal the more subtle modulation of the MMN by contextual changes (namely, the predictability of the sound sequence), and (iii) to identify the most likely learning process that explains the MMN response.
Main results: Our results show that ASR and EMD are both able to reveal a significant MMN and its modulation by predictability, and even appear more sensitive than the offline analysis when comparing alternative models of perception underlying auditory evoked responses.
Significance: ASR and EMD show many advantages when compared to other online artifact correction methods. Besides, subtle modulation analysis of the MMN, embedded in perception computational models is a novel method for assessing the quality of artifact correction methods.