Poster C84, Wednesday, August 21, 2019, 10:45 am – 12:30 pm, Restaurant Hall
Removal of Ocular Artifacts from Magnetoencephalographic Natural Reading Data: Recommended Methods
Sasu Mäkelä1, Jan Kujala1,2, Riitta Salmelin1;1Aalto University, 2University of Jyväskylä, Finland
Introduction Reading is one of the most important forms of human communication. For understanding the cortical basis of reading, it seems essential to use natural reading paradigms in conjunction with a brain imaging modality that has both high spatial and temporal resolution, such as magnetoencephalography (MEG). However, due to its electrophysiological nature, MEG is sensitive to electromagnetic artifacts, of which especially problematic for reading studies are ocular artifacts. In this study we present two different methodological pathways for removing ocular artifacts from MEG reading measurements, with the focus on saccades and blinks. The aim is to demonstrate the effectiveness of the approaches in removing ocular artefacts from continuous reading data and to evaluate the possible benefits of using a multi-stage process. Methods Both of our alternatives are based on blind source separation methods, but differ fundamentally in their approach. The first alternative is a multi-stage process consisting of two methods presumably well-suited for removing either saccades or blinks. In the first part saccades are extracted by applying Second-Order Blind Identification (SOBI) which exploits the temporally coherent patterns that saccades form in normal reading situations. In the second part an independent component analysis (ICA) method called FastICA is used to extract blinks, which appear in the measurements as independent random deviations. Our second alternative is to use only one method for removing both artifact types. For this we used Adaptive Mixture ICA (AMICA), a method shown to outperform most competitors (Delorme et al., 2012, PloS one, 7(2), e30135). The alternatives were tested on MEG data recorded from 10 subjects in a natural reading task. For both methods we evaluated whether and how clearly they were able to identify artifacts from the data. In order to compare the spatial similarity of the artifact components yielded by the different methods, Pearson correlations of the components' sensor weights were calculated for all subjects. Results Saccades were extracted by both SOBI and AMICA from all subjects with a single component. These components were highly similar in 9 subjects, with SOBI/AMICA correlations in the range 0.954–0.994 (0.753 for one subject). Blinks were extracted with a single component from all subjects by AMICA, and from 8 subjects by FastICA. For one subject, FastICA decomposed the blinks into two components, and for a second subject the method had to be run with only a partial dataset in order to find a blink component. For the 8 completely successful runs, SOBI/AMICA correlations were 0.885–0.991 (0.825 and 0.507 for the remaining runs). Conclusion Based on these results the two pipelines produce very similar results on saccade and blink artifacts, except for the two less successful performances with FastICA. Discounting these instances, both alternatives seem equally recommendable. When the different pipelines yield noticeably different results, simulated data in which the true distribution of artifacts is exactly known is needed to determine which alternative extracts the artifacts better.
Themes: Reading, Computational Approaches
Method: Electrophysiology (MEG/EEG/ECOG)