Evaluating Automatic Artifact correction for Online Hypothesis Testing

In BCI, artifact removal remains an acute challenge. Filtering must be efficient in removing artifacts while preserving relevant features, e.g. eventrelated potentials (ERP) like the mismatch negativity (MMN). MMN is a prediction error signal whose modulations reflect human perceptual inference and learning. Characterizing these subtle processes requires fitting non-linear models onto single-trial data. And disentangling between alternative models is challenging because of a low signal-to-noise ratio. We evaluated four methods for online artifact removal. We mimicked online data processing using real electroencephalography (EEG) data from an auditory oddball paradigm. We compared the four approaches with standard offline analysis, in their ability to reveal (i) the MMN, (ii) the MMN modulations by the manipulation of the predictability of a sound sequence and (iii) the most likely learning mechanism at play. Artifact Subspace Reconstruction (ASR) and Empirical Mode Decomposition (EMD) were the most successful. Interestingly, they even proved more sensitive than the offline analysis, likely because they avoid rejecting trials.