Backproject clustered ICs
Note: In this page I call 'cluster' in the sense of 'independent component clusters created by EEGLAB STUDY'. Also, IC stands for independent component.
What is std_backproj
The backprojection means the forward mixing process from EEG sources (i.e. ICs) to scalp channels. std_backproj is recommended for
- Quickly showing cluster-to-channel backprojection at the group level in the form of ERP and variance accounted.
- Creating another set of .set files that have only ICs included in the selected cluster(s).
It comes with a powerful (for EEGLAB) interactive visualization tools too.
Note that the envelope plots and percent variance accounted for (PVAF) calculated here are different from those produced by std_envtopo(). The difference derives from that std_backproj() computes envelopes and PVAF across subjects, while std_envtopo() computes them across clusters.
Note also that PVAF is subadditive, namely PVAF(a)+PVAF(b)+PVAF(c)>=PVAF(a+b+c) because (a+b+c) causes cancellation.
- As a visualization tool for the group-level source-to-channel backprojection
std_backproj provides the quickest pathway to visualize the group-level IC clustering results backprojected to a single-channel ERP/variance accounted. std_backproj() performs the following operations
- Identifies which ICs are included acorss the selected clusters (by using the admittedly 'apparently complex scheme' by the developer)
- Open each selected dataset, reject all ICs that are not included in the selected cluster(s).
- Compute ERP/variance accounted.
- Repeat above 2 and 3 for all the datasets.
- Compute the grand-average ERP/variance accounted across all datasets.
- Visualize the results.
- As a group-level filter to manually exclude non-EEG ICs for SIFT and Measure Projection
One may use it as a group-level filter. Creating STUDY is to clean the data since it provides two powerful filters
- To kick out ICs with dipoles located outside the brain
- To kick out ICs whose dipoles' residual variance is larger than a certain threshold (default 15%) This usually removes 70-80% of ICs. Nonetheless, quite a few number of artifactual ICs survives these criteria. I show example below. This is the spectrum plot of STUDY-level IC clusters. One can easily notice non-EEG spectrum patterns (i.e. non-1/f patterns) in Cluster 5, 6, 12, 18, 37, and 40 (highlighted with red crosses). Total of 83/1066 ICs (7.8%) in the final clustering results are identified as non-EEG sources.
To address this issue, one may use std_backproj in the following way:
- Create STUDY and cluster ICs as usual, but only precompute spectra.
- Set the final cluster number to be large (e.g. 40).
- Identify clusters that shows apparent EMG or whatever non-EEG pattern (i.e. non 1/f curve) in the spectra plot.
- Select the remaining good clusters to backproject to create whole another set of datasets with good ICs.
The finally created super-clean datasets are perfect for SIFT (now all ICs are clean and usable!), Measure Projection (it does have eyeCatch to exclude EOG ICs, but does not have a solution to exclude EMG, so EMG ICs should be better removed beforehand), and so on.
I performed validation to make sure that nothing wrong should happen in figuring out the labyrinth of STUDY.
Materials and Methods
- Load a STUDY with 17 IC Clusters, select Cluster 3,9,14,17 for inclusion and 2,6 for exclusion, start backprojection. Save the new .set files with the selected ICs to the folder X.
- Show Fz ERP obtained from std_backproj.
- Clear the STUDY. Next, load all the .set files from the folder X. Compute grand average of Fz ERP for comparison. Compare the two ERP plots in the figure below.
- The effect of ascii conversion was in the order of 10^-9 (see the top plot below).
- The difference between the two results were in the order of 10^-4 (see the middle plot below).
- However, the main difference was in DC part that was 3.7853e-04. The difference in AC part was in the order of 10^-7 (see the bottom plot below_.
- The SNR in the AC part is at least 120dB (ERP wave ranges is about 1; AC difference in the order of 0.0000001)
Conclusion The method fulfills the precision in practice. I don't know exactly where the numerical difference in both DC and AC come from.
Authors: Makoto Miyakoshi. SCCN, INC, UCSD