First of all, I’m new to the community but have been using MNE for EEG processing for the past few months. The tool has been fantastic! Anyway, I’m seeking some advice for how to pre-process the 16-channel EEG data my group has collected for post-hoc spectral analysis. To give a better context, the 3-hour recording’s session involves studying hot flashes in menopause. These events occur spontaneously and often involve dramatic sweating events (even on the scalp). My current plan for pre-processing involves:
Removing bad channels
Common-average reference
High-pass @ 1 Hz and notch-filtering @ 60 Hz
Manually annotating for “bad” segments (noisy bursts and filtering artifacts)
Fit ICA on “non-bad” segments
Apply ICA
After that is applying multi-taper and analyzing the time-frequency representations in various ways. What I have noticed while doing this is that, the sweating events are dramatic, and I think naive default filtering with the default FIR filter causes leakage to the delta band, evident by the appearance of the very sinusoidal ~1 Hz waves (also increase in power in the time-frequency representation). Because of this, I’m avoiding analyzing in the delta band for now.
Additionally, my understanding of ICA is that it assumes the mixing matrix is relatively time-invariant and that the transformation between the sources and channels’ signals are linear. However, am I correct to think that sweating events can lead to changes in the electrode-skin impedance in a unpredictable way for each electrode and, therefore, violate this assumption of time-invariance? Thus, though I have been applying ICA on the entire 3-hour session, I think it would be better for me to fit the ICA separately for the pre- and post-sweating events. However, what I’m interested in studying involves analyzing what happens during the events as well, and I’m not sure how to deal with this. I have come across some papers regarding adaptive ICA. Has anyone tried such techniques out and how effective were they in dealing with artifacts removal in your case?
First of all, I’m also fairly new to EEG, so please forgive me if I’m not precise with some terms.
I think your question ultimately comes down to how to handle dynamic artifacts, and I agree that sweating is a big challenge in that space.
A thought that came to my mind: if there are sensors that can directly monitor sweating or skin impedance changes during recording, would it be viable to use a multi-modal approach—combining that information with the EEG signal—to better correct for these artifacts? I’m curious whether this has been explored in your context, or if you’ve considered it.
Hi, thanks for the thoughtful response. We indeed are approaching the topic of interest with multi-modal sensing and have skin conductance sensors peripherally. I think the core idea, if this was to be implemented, is, given that we have a known profile of sweating elsewhere, can we fit/predict the response of the artifacts at EEG channel locations so that we can regress out. However, I think it’s a very non-trivial task for several reasons.
Peripheral sweating profiles across body locations (slower/faster, sharper/more dull, etc.). Plus, the core issue is that it’s the low(er)-frequency drifts across all channels due to the sudden increase in conductance across all channels. However, because of the lack of ability to control for electrode-skin impedance across all electrodes and the differing in amount of sweats across different sites, some electrodes become more conductive than the others in unpredictable manners, though all of this is in reference to some reference I choose for post-processing (i.e. common-average). This means some channels drop and others rise. Plus, the skin conductance sensor measures skin conductance by measuring electrical responses while applying some active currents (AC/DC w/ controlled either voltage or current), whereas the EEG ones, I believe, are passive. Overall, I think the transformation from the skin conductance device being used to the EEG drifts on the scalp would be very non-linear and unpredictable, and also require some ways to obtain some information on the impedance of the electrode-skin electrodes during those times. But… if there’s some clever machine learning techniques for these scenarios, I’m all ears!
Hi, thanks again for the detailed and thoughtful explanation. I totally get why it’s so hard to just use skin conductance data to predict and remove those EEG drifts—the relationship just seems too unpredictable across channels and time.
That’s actually why I brought up that new electrode paper: https://advanced.onlinelibrary.wiley.com/doi/abs/10.1002/adma.202523019 . Instead of trying to fix the problem after it happens with a clever algorithm, maybe we can take a more direct approach: use hardware that keeps the electrode-skin conductivity stable from the very beginning. The idea is to prevent most of the sweat-related drift from happening in the first place, rather than relying on a complex model to clean it up later.
Of course, this doesn’t mean we should abandon the regression or machine learning methods. But if the electrodes themselves can make the interface much more stable, maybe those algorithms wouldn’t have to work as hard, and the whole system could become simpler and more robust.
Just my two cents. Would be curious to hear your thoughts on this kind of hardware-first approach.