Objective: Electrocorticographic (ECoG) signals contain noise that is common to all channels and noise that is specific to individual channels. Most published ECoG studies use common average reference (CAR) spatial filters to remove common noise, but CAR filters may introduce channel-specific noise into other channels. To address this concern, scientists often remove artifactual channels prior to data analysis. However, removing these channels depends on expert-based labeling and may also discard useful data. Thus, the effects of spatial filtering and artifacts on ECoG signals have been largely unknown. This study aims to quantify these effects and thereby address this gap in knowledge.
Approach: In this study, we address these issues by exploring the effects of application of two types of unsupervised spatial filters and three methods of detecting signal artifacts using a large ECoG data set (20 subjects, four task conditions in each subject).
Main results: Our results confirm that spatial filtering improves performance, i.e., it reduces ECoG signal variance that is not related to the task. They also show that removing artifactual channels automatically (using quantitatively defined rejection criteria) or manually (using expert opinion) does not increase the total amount of task-related information, but does avoid potential contamination from one or more noisy channels. Finally, applying a novel 'median average reference' filter does not require the elimination of artifactual channels prior to spatial filtering and still mitigates the influence of channels with channel-specific noise. Thus, it allows the investigator to retain more potentially useful task-related data.
Significance: In summary, our results show that appropriately designed spatial filters that account for both common noise and channel-specific noise greatly improve the quality of ECoG signal analyses, and that artifacts in only a single channel can result in profound and undesired effects on all other channels.