Independent components analysis for removing artifacts

[@michaelrecine7626]

This is definitely the most comprehensive explanation I've found of this topic yet! I really appreciate the level of detail as well as the discussion of your philosophy behind it. Please don't ever take this down lol

[@mikexcohen1]

Thank you! The content of these old videos is fine, but I'm a little ashamed of the horrible audio quality :P

[@triplevs4432]

This video is so helpful! It definitely helps me understand how to determine which component to remove in EEG data analysis better! Thanks! Thumb up!

[@ananddeshmukh4939]

Thanks For Info. mostly I cleared my all doubts about ica and preprocessing.

[@jesusdanielolivaresfiguero4752]

I see that in the videos you use millisecond signal segments for ICA. In my case, I need to analyze data of longer duration, between 1 and 2 seconds. The sampling frequency that I handle is 128, so each channel would have between 128 and 256 points for 1 and 2 seconds respectively. In my project I have to detect stress in airplane pilots while they complete race tracks with different levels of difficulty. From my little experience, I observe that it is not an ERP. What I plan to do is apply a bandpass filter and ICA to remove artifacts. The question is: - Do you consider that it is a long time to apply windows of 1 or 2 seconds to apply ICA? - Do you think 128 or 256 points per channel is enough? I hope you can answer or recommend literature or videos about it. Thank you very much for your help, I have learned a lot in a short time thanks to you.

[@mikexcohen1]

Tough to say. A general guideline for computing the amount of data you need is the formula kN^2, where N is the number of channels and k is some scalar (that formula is from the eeglab website). For noisy data, k needs to be larger, e.g., 30. If the data are clean, k can be small, e.g., 3 or 5. You can also try filtering the data, e.g., 2-20 Hz, which should help the ICA decomposition.

[@jesusdanielolivaresfiguero4752]

@@mikexcohen1 Thank you very much for the tip.

[@tungekarimad]

Thanks for the lecture. It cleared many of my doubts. How can we save the artifact removed EEG data after the pre-processing step?

[@mikexcohen1]

Hi Imad. If you are using eeglab, then you can save the original IC weight matrix and the original EEG data. That will allow you to reconstruct the artifact even after removing it from the data.

[@lenabruemmer4702]

Hi mike, Thanks a lot for your video! In my Ics there are some with peaks in the power spectrum at around 12, 25 and 37 Hz what does that mean? a lot of alpha Oszillation in my data due to a peak at around 12 Hz in the power spectrum? Thanks a lot for your answer!

[@xizh1750]

Hi Mike, I did the steps of ICA removal as you did. I followed your ideas about artifacts removal. Since I'm writing my master Thesis, I have to mention the sources I used while removing the IC's. So would you recommend any written reference for this? Either written by you or by anyone else. Thanks in advance for help.

[@mikexcohen1]

Hi Xi. Many people cite one or both of the Jung 1998 papers that you can get an overview of here: sccn.ucsd.edu/~jung/Site/EEG_artifact_removal.html Mike

[@amitbishnoi43]

Thanks for the lecture Mike. Currently I'm working on a Deep Learning project with EEG dataset which is stored in .csv format. Although I converted it into .edf format using EDFBrowser to import into EEGLab, but now after component rejection how can I convert the data back into .csv from .fdt & .set formats, as you know, both of these are output formats of EEGLab. Please help. Thanks.

[@mikexcohen1]

Hi Amit. csv format might be best done through MATLAB directly. For the eeglab formats, I guess it should be possible via the GUI. I use eeglab a bit because I like their data-storage format, but I'm not an eeglab expert. You should probably post your question to the eeglab email list, or see if this information is available on their online tutorials.

[@amitbishnoi43]

@@mikexcohen1 Thanks for your input and suggestions Mike. I'll contact EEGLab for further clarification.

[@thomasguieysse3351]

Hi! Thanks a lot for your lectures and especially this one, very well explained! I'm working on an experiment where only 6 EEG channels are used to acquire the data. I wanted to apply ICA in order to detect and remove artefacts, but a collegue told me that it's not recommended to use it with only 6 components. There are no many articles about it online, so my question is: Should I use PCA instead, or is it possible in some cases to use ICA with few components? Thanks a lot for your answer

[@mikexcohen1]

Hi Thomas. You can definitely run ICA with six channels -- you really need only two channels. The question is whether you will get good results with cleanly isolated components. That depends on what kind of signals you have. I'd say try it and see. Blink artifacts will probably be isolated if you have anterior channels. PCA is not a good idea for artifact separation, regardless of the number of channels you have. If you have more specific questions, feel free to post them on discuss.sincxpress.com.

[@thomasguieysse3351]

@@mikexcohen1 thanks a lot!

[@superneuronita]

Thank you very much Mike! One question: I would like to know how to do this process that I read in an article "we removed components that correlated with at least one EOG channel (r > .40 Flexer, et al, 2005) "; do you know how to do this in EEGLab?

[@mikexcohen1]

I don't think that's native in eeglab. Probably best to check out that Flexer 2005 paper.

[@miou97two]

Thank you so much for your videos. Where can i find the whole course, please? I can't find it... Méyi, PhD student

[@mikexcohen1]

Hi Méyi. You can look on my main youtube channel for all of the ANTS# playlists. I also have all of my courses linked from my website sincxpress.com/

[@YuZhang-j6f]

Thank you!!! Very helpful to me. You are awesome 😄👍

[@mikexcohen1]

No no, *you* are awesome :)

[@javiercastillo6323]

What plugins did you use? I'm using the latest version of EEGLAB and we don't have the same tools for artifact removal.

[@mikexcohen1]

I think I used an older version, maybe v6. The same functionality should be available, but the GUI might look different, e.g., options in different places. Sorry for the confusion :P

[@jaxejaxejaxe]

Dear Mike, Conceptually ICA and FFT seems similar. Both trying to find the original sources of the signal. Can you help me understand this better?

[@mikexcohen1]

hmm, well, both methods are filters (ICA is spatial and FFT is temporal), and both can be used to attempt source separation, but they're quite different analytically.

[@uttarakhatri908]

Dr. Mike, Thank you for the tutorial. I have a query: I ran ICA and I am unable to identify the EOG component because of the absence of topographical maps, any guidance as to how to proceed?

[@mikexcohen1]

You can look at the components time series. The EOG component is easy to see, because it's mostly flat and with really large spike-like deflections each time the person blinked. Horizontal EOG artifact has positive/negative step deflections, as people look to the left or right.

[@maryamnoroozi5448]

I cannot find the pre ICA video?

[@juliomedeiros146]

Thanks for Info, great work!

[@observer698]

3:18 I'm from the USA but pronouncing IKA makes more sense! The word "component" is read with K sound and also when there is a, or o, u C is read as K in USA too! :) so I think reading ICA as ISA doesn't make sense :)

[@mikexcohen1]

lol, you must be a linguist. I hadn't thought about it, but you're right -- it would be natural to pronounce it eeeesssaaaaahhh.

[@observer698]

@@mikexcohen1 A linguist! :))) I'm just a mom who happened to watch one of the pbs kids' shows with her kid and learned this piece of information :)

[@mikexcohen1]

Well, if you're considering a career change...

[@linaelsherif2577]

Hello mike currently I am using IC label to help me decide which components to reject. I have two questions: 1. from 64 IC how much if I reject I must discard this participant data. 2.after selecting the IC to be rejected in future how can I save them as you did?

[@mikexcohen1]

It's impossible to say how many ICs you should remove, but I advise removing as few as possible. It's very unlikely that an IC will contain *only* noise, so each IC you remove will also remove real brain signal. In a clean EEG dataset you might remove 1-2 components.

[@LudwigVera]

Thanks for the very helpful video. I am just wondering: might it not be problematic to rely on whether or not an artifact is event-related? Because I might also have muscle twitching or a strong blink reaction whenever the event/experimental condition happens, so this criterion does not warrant that the signal is coming from the brain.

[@mikexcohen1]

Yes, those are problematic cases, and it's really difficult to separate, e.g., blinks when the research subject blinks every time the stimulus appears. The best way to deal with these time-locked artifacts is to avoid them in the first place. We do that by explaining the artifacts to the subjects before data collection, so they can help give us clean data.

[@LudwigVera]

@@mikexcohen1 thanks, this is helpful!

[@amarimuthu]

Thanks very much for your nice presentation Can Connectivity measures are useful when we calculate them on the Independent components post-ICA? I have this question, as the ICA removes the correlation at first before calculating the independent components (whitening process). Or We can apply these measures only on the raw measurements. Thanks

[@mikexcohen1]

Hi. You can do connectivity on IC time series. The independent components are independent in the linear algebra sense, meaning not collinear (one vector is not a scalar combination of other vectors). Two ICs can be correlated. Furthermore, the decomposition is based on the entire dataset, whereas neural synchronization tends to be relatively transient. On the other hand, one can wonder how exactly to interpret connectivity between two ICs -- they are by definition *statistically* independent time series, but that doesn't mean they correspond to discrete localized brain regions or dipoles. I'm not arguing that IC connectivity is invalid, more that connectivity in general with non-invasive electrophysiology should be done carefully and with critical thought about how to interpret the results. Hope that helps Mike

[@amarimuthu]

Thanks very much, Mike for your explanation. I will be curious to check that.

[@amarimuthu]

@@mikexcohen1 Dear Mike, I had checked on few aspects once the Independent Components are estimated. I also estimated the Mutual Information (MI) shared at Channels level and at Components level. I was surprised to see that the Mutual information shared at Channels level is lower than at the components level (Mean value). Is that possible to happen? I read the disadvantages of MI measure, however, this part is something against my intuition. Also, the Spearman correlation distribution is widespread at components level than at the channels levels. Is that possible as well? Thanks.

[@haneensuradi]

Hi Mike, I was reading on how to decide on IC to remove, and in some references they mention that most of the last components (beyond half the number of the channels) cannot be brain components and are usually recommended to be removed. Do you think this is a good approach? Or is it better to rather keep those components which don't look like a brain component but also not a known artifact (heart, eye, muscle or line)?

[@mikexcohen1]

I recommend keeping as many components as possible. ICA doesn't know anything about "signal" and "noise"; it's just looking for mixtures of channels that produce non-Gaussian distributions. The brain does not perfectly conform to the assumptions of ICA, and so real brain sources will distribute into many ICs.

[@pulutogo8266]

Hi Mike, currently I am using ICA to remove EOG artifact. When I do ICA without EOG electrodes, the decomposition is better. So, my question is except for comparing the EOG signal and EOG components derived by ICA. What are the additional roles of recording EOG, may I stop recording the EOG channels in the following experiment?

[@mikexcohen1]

I noticed the same thing a long time ago and stopped recording EOG. As long as you have a lot of electrodes (>50) and decent coverage on anterior scalp regions, the ICA will be able to resolve a nice blink component without EOG. And EOG has strong disadvantages -- they are uncomfortable and distracting for the participants, which means they pose a small-but-real risk to decreasing task attention and performance.

[@pulutogo8266]

@@mikexcohen1 Thank you so much for answering my question. You are a great teacher.

[@mikexcohen1]

aww #blushing

[@bluehope4143]

Is it possible to use EEG LAB ICA for local field potentials?

[@mikexcohen1]

Yes, definitely. eeglab is for any ephys signal.

[@tv.elghazali]

i want to ask about iCa corp for detect anomly target

[@sulagnaroy3192]

can we conduct this without using eeglab?

[@mikexcohen1]

Yeah, technically you could. It would be quite some extra work to reproduce the visualizations. Many other analysis software packages allow for similar analysis methods.

[@subrisubrika5652]

Great video I just wish you weren't breathing so heavily in the microphone. Maybe you should have used ica to remove them :D

[@mikexcohen1]

Yeah, sorry about the shitty AV quality. This was my first ever attempts to do recordings, and I'm mildly embarrassed about it :P The good news is that next year I'm planning on re-recording this entire video series, with much better slides, better home-recording setup, better everything. So you have something to look forward to!

[@subrisubrika5652]

@@mikexcohen1 I don't think I can wait that long! This was very informative and helpful so thanks! I'm just curious do you think ICA would be useful in an experiment where eye movements are constant, but predictable? We are doing an experiment where participants view faces and they are directed to look from one facial feature to the next (left eye to nose to mouth) all while recording EEG and eye tracking data. I'm just worried that the data will be mired with eye saccade contamination and unanalyzable.

[@mikexcohen1]

It's certainly worth a try. If the characteristics of the eye movements are the same, then ICA should be able to isolate a component (or a few).

[@dadihoussem8102]

barakalahou fik

[@mikexcohen1]

wa fik al barakah (I looked that up online, hope it's correct!)

[@tv.elghazali]

i want to ask you about ICa corps