Adaptive ICA for Speech EEG Artifact Removal

Recently, in cognitive neuroscience, artificial stimuli such as single sentences have been replaced by more naturalistic stimuli such as continuous speech. Since it is already known ‘where’ in the brain language is processed, the next crucial step is to investigate ‘how’ these neuronal circuits and processes work. Thus, it is necessary to apply experimental procedures with a high temporal resolution such as electroencephalography (EEG) in order to capture and identify these mechanisms. However, EEG is highly prone to measurement errors as the surface electrodes collect all kinds of electromagnetic noise from physiological and non-physiological sources. Here, we present a procedure to remove those artifacts (with a special focus on eye artifacts) and provide evidence that it is possible to extract event related potentials (ERPs) from EEG data recorded during listening of an audio book. We developed an evaluation pipeline, tested on EEG-data of 36 participants. The pipeline consists of two major steps: spectral filtering (bandpass: 1 Hz-20 Hz) and a custom version of independent component analysis (ICA) filtering. Thus, we defined one channel (Fp1) as our electro-oculogram channel (EOG) and tested which independent components significantly correlate with this channel. All independent components that had a correlation above a fixed threshold were removed. This procedure is highly reproducible and allows to extract clean ERPs from EEG data during continuous speech perception. We show that the ERP responses to adjectives are different from ERPs to verbs in shape as well as latency. We suggest that this advancement in evaluating EEG data may further improve neurolinguistics research and is a further step to develop a unified evaluation pipeline for these kinds of data.