I’m new to MNE, and I have a question about filtering raw data before applying ICA.
I understand (thanks to the tutorial: Repairing artifacts with ICA — MNE 1.7.0 documentation) that an important step in ICA is to use a filter to eliminate low-frequency drift.
A high-pass filter with a cut-off frequency of 1 Hz is recommended for slow drift. What’s more, since filtering is a linear operation, we can apply the ICA solution to the unfiltered signal afterwards. There will be no consequences for the rest, depending on the filter?
But what about the low-pass filter? When pre-processing the EEG, we applied a high-pass filter at 0.1 Hz and a low-pass filter at 40 Hz due to the presence of power lines at 50 Hz.
Can we use the ICA method on data from 1Hz to 40Hz?
Is it better to take electrical noise into account in ICA with this filter at 40Hz?
If not, and the 1Hz high-pass filter is sufficient, can you explain why?
I’d recommend applying the same low-pass filter to both datasets.
That is, in your specific case, I’d filter the data used for ICA fitting between 1 and 40 Hz; and the data to be cleaned between 0.1 and 40 Hz.
I would suspect that leaving high frequency noise in the data used for ICA fitting will use ICA components to capture line noise and some muscle artifacts – stuff you won’t have in the data you’re planning to analyze. So I’d avoid that.
You don’t have to train and apply ICA to data with identical high-pass filters. It is strongly recommended to train ICA with data that has been high-pass filtered at 1Hz or even 2Hz, but you can then apply that ICA solution to the same or other data sets that have different or no high-pass filters at all.
Regarding the low-pass filter, I don’t think that it is critical to apply a low-pass filter before training the ICA. Usually, I don’t apply a low-pass at all, because filters can have dramatic effects on large spikes or other artifacts, and I’ve found ICA to work quite well without any low-pass filters (of course you already have a natural low-pass filter that was set during recording with a cutoff frequency at the Nyquist frequency or usually much lower).
The EEGLAB tutorial has some nice suggestions on how to optimize ICA quality, and they only talk about how the high-pass filter is very important and don’t even mention the low-pass filter: d. Indep. Comp. Analysis - EEGLAB Wiki