ICA decomposition
During the ICA procedure, the sensor signals are decomposed into potentially independent sources. In this step, we switch from the sensor signals (measured by electrodes, magnetometers, or gradiometers) to the representation of the original recording by the set of independent components (ICs). Each IC is characterized by its signal time-course and scalp topography and can represent a potentially independent source of brain activity.
The rationale for decomposing EEG/MEG recordings with ICA is twofold. First, it allows for a data-driven separation (and rejection) of artifacts. Secondly, separated components of brain activity can be more efficiently localized with available localization methods. These key points highlight two phases in the ASCT where ICA decomposition is carried out.
The first decomposition (ICA1) allows for removing artifacts that are still present in the signal after sensor-space artifact rejection. This is accomplished by analyzing the excess signal variance in IC timecourses. Additional removal of noisy channels is also performed by detection of ICs that contribute to a single sensor only. This means a very steep gradient around a single channel on topographical maps.
After the ICA1-based artifact removal, ICA2 can provide a finer decomposition as it runs on relatively clean signals. The ICA2 step is completed with the classification of the resulting ICs into those of brain and non-brain origin. Only the former ones, representing cortical activity, are retained for further analysis.
ICA settings
ICA settings start with the decomposition mode. ica.approach can be set to either 'defl' (deflation), which limits the number of resulting ICs, vs. 'symm' (symmetrical) approach, where the number of ICs is equal to the number of channels. By default, ASCT uses the deflation approach, which is highly recommended.
The depth of deflation (i.e., the extent to which the outcome dimensionality is limited) can be adjusted with ica.ICAx_numSteps, which affects the number of resulting components ('x' in the parameter name should be replaced by the relevant ICA number, either '1' or '2'). Higher values tell the ICA to more strongly seek out possible ICs in the portion of data not yet explained.
Recommended settings differ for ICA1 and ICA2. As the role of ICA1 is artifact detection, this can be practically accomplished with less accurate but faster decomposition. This can be obtained with 100-150 steps for ICA1, which should yield around 10-20 resulting ICs.
On the other hand, decomposition accuracy is crucial for ICA2, which is used for further brain signal reconstruction. Hence, the ica.ICA2_numSteps parameter is typically set to 200-250 steps. It can be empirically adjusted to eventually obtain around 25-50 ICs.
Moreover, as ICA is a non-deterministic procedure, its outcome will differ every time it is run. Therefore, to obtain an optimal separation, we recommend launching ICA multiple times (as defined by ica.ICAx_numIter). Three runs for ICA1 are usually sufficient for efficient AR, while ICA2, which is the actual decomposition used for further signal reconstruction, requires around 20–40 iterations. The selection of the best ICA2 decomposition is performed in step 6 after classification of resulting ICs.
The ica.ICAx_nonlinearity allows for setting the nonlinear function ('tanh' is recommended).
Saving ICA files
ICA1 results are saved in the 4_ICA1 subdirectory. Within this directory, files with the _ICA1decomp.mat extension are saved, which contain decomposition results for datasets. Moreover, in the subdirectories named with dataset names, a set of images, for each iteration separately, is saved that depict IC topographies, timecourses, spectral plots, and evoked responses.
ICA2 results are saved in the 5_ICA2 subdirectory with a similar structure of files.
Last updated