Independent Component Analysis (ICA) | EEG Analysis Example Code

[@hanstschohl]

Ive had a class on Neurotech and they did not explain any of this but expected me to write a program using ICA feature selection. I found nothing usefull until this video. thank you so much

[@ShawhinTalebi]

I was in a similar situation. That’s a big reason for this video. Glad it was helpful!

[@golnazbaghdadi4617]

The best explanation of ICA that I have ever seen, thanks!

[@rlee1257]

Very clear, thanks! The examples with the mic and blinks were a great inclusion imo. They made ICA much easier to understand while also displaying the practical application in a fun way :)

[@ShawhinTalebi]

Glad it was helpful!

[@thevitorialima]

Amazing explanation man, I always end up looking your videos for technical but clean explanations!!

[@ShawhinTalebi]

Thanks Vitoria :)

[@NonstopElectroshock]

Amazing explenation! Thank you so much! You saved my butt for my exam tomorrow

[@ShawhinTalebi]

Glad I could help!

[@renanbarella7556]

Gotcha! it was an amazing explanation about ICA! Thank you very much for that Shawhin! 🙃

[@ShawhinTalebi]

Thanks, glad it made sense 😁

[@raven5847]

Love it! It was well explained and the example part makes the concept clearer. Thank you : )

[@ShawhinTalebi]

So glad it was helpful!

[@mahdieshraghi]

Thank you from Iran 🇮🇷

[@ShawhinTalebi]

Happy to share!

[@neurodiscussionswithpegah]

Great explanation! Thank you Shawhin!

[@ShawhinTalebi]

Glad it was helpful!

[@peteremmanuel8557]

Beautiful explanation Shaw... Thanks

[@barmawn]

Thank you for your useful contents shawhin🙏🏻

[@ShawhinTalebi]

My pleasure!

[@karenfu4242]

good examples and codes for beginner like me to learn the concept of ICA!

[@ShawhinTalebi]

Glad it was helpful!

[@woodworkingaspirations1720]

Back to watch again

[@martonmunding8645]

Great explanation, thank you! I was looking for some material to help me with an EEG blink removal problem, so it's the perfect find!

[@ShawhinTalebi]

Thank you for your comment. That is great to hear. Happy to help!

[@ifycadeau]

woohoo! been waiting for this one

[@ShawhinTalebi]

Happy to oblige

[@kaldirYT]

My birthday is now complete

[@chiragpalan9780]

thanks for explanation, but may I ask why we want independent components to be non-gaussian?

[@ShawhinTalebi]

Hi Chirag, that's a good question. I go into a bit more detail in the blog: towardsdatascience.com/independent-component-analysis-ica-a3eba0ccec35. Briefly, however, ICA combines the non-Gaussianity of independent components assumption, with the central limit theorem to connect the measured signals (e.g. x1 and x2) to the unmeasured independent components (e.g. s1 and s2). Hope that helps! Please feel free to share any follow-up questions, I find them very helpful. Shaw

[@dragoncurveenthusiast]

This was super useful! Thank you!

[@ShawhinTalebi]

Glad it was helpful!

[@Jamboreeni]

Great vid man!

[@ShawhinTalebi]

Thanks Abhay! Glad it was helpful

[@parinazamini2072]

Thank you so much , this video is really helpful 🌷🌷

[@ShawhinTalebi]

Glad it was helpful! Thanks for watching

[@muzammiliqbal4620]

Thank you. Very informative session.

[@ShawhinTalebi]

Thank you, happy to help!

[@aminsinichi]

Great job!

[@ShawhinTalebi]

Thanks for the visit!

[@MrSreior]

Thank you so mucj

[@ShawhinTalebi]

Glad it was helpful!

[@Hhushrk12]

Hi, what PCA data specifically did you put into ICA? Did you take the variable loadings of each of the 21 PCs? Was that your data frame? I am trying to do this for financial data, would be awesome if you took the output of your PCA video and ran ICA on top of that

[@ShawhinTalebi]

Hey, thanks for the question! The transformed EEG signals (i.e. 21 principal components) are what are fed into the ICA function. The data structure is an array where rows correspond to time points and columns to each of the principal components. That sounds like a cool idea! It should be simple to implement. If you are using sklearn, you can use a X_pca = pca.fit_transform(X) to get the transformed data, then a X_ica = ica.fit_transform(X_pca) to apply ICA to the principal components. PCA: scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html#sklearn.decomposition.PCA ICA: scikit-learn.org/stable/modules/generated/sklearn.decomposition.FastICA.html Hope that helps!

[@QifanWang]

THX! Very clear! Could you explain SCA vs PCA? Shared component analysis or Shared variance component analysis?

[@ShawhinTalebi]

Great suggestion! I haven't used SCA before, but I just downloaded the paper :)

[@saadawan477]

Hi, Shawhin can you please tell me that how I can made heuristic guess in case of grayscale imges? let say i get PCA_data i.e 10 Principal Components now then how many Independent Components i need, in order to construct gray level co occurance Matrix? while i am doing MR Brain images segmentation and classification in matlab. Thanking you in anticipation.

[@ShawhinTalebi]

Hi Saad, thanks for the question. That sounds like an interesting project! A simple thing is to just use the explained variance from PCA (say >0.95) to determine the number of components, then keep the same component count for ICA. I'm sure there are more sophisticated tricks you can do based on the ultimate goal for the project. Hope that's helpful!

[@saadawan477]

@@ShawhinTalebi Thanks for your reply. yeah i got it, i did it in the same way i.e No. of PCs= No. of ICs. Yes, it is helpful!

[@alexeigilev8577]

Is it possible (or does it make any sense) to apply ICA to participants ratings (behavioural data) on many variables?

[@ShawhinTalebi]

Thanks for the question! Typically ICA requires time series data. If these ratings are indexed by time, ICA should work. Hope that helps!

[@woodworkingaspirations1720]

Well explained

[@ShawhinTalebi]

Thanks, glad it was clear!

[@reza.partovi]

It was a great video. To restore back the data, shouldn't we add the mean also? After doing the inverse pca tranformation.

[@ShawhinTalebi]

Hi Reza, thanks for your comment! Yes that is a good point. In general, we also need to invert the normalization step (hidden in MATLAB's pca function) to get back to our original units. For EEG data, we may not care about this step since we are interested in signal oscillations (i.e. brain waves) rather than absolute voltage values. Thanks again, Shaw

[@reza.partovi]

Thank you Shaw

[@manden_fruta]

Hi Shawhin, thank you for the clear explanation about ICA I tried your code without any modifications and every time I ran the script, the results are different. Like everytime the resulting signal without blink is different in shape and amplitude, is that normal? Also, when I plot the original data (64 channels) vs the resulting data_noblink, each channel is not consistent regarding its positions (like data from channel 1 (Data(:,1)) is now in Data(:,6) and so on. My question is, is there a random effect of using ICA that I didn't catch when inverse PCA transform happens? if so, can I use a seed or similar to control this randomness?

[@ShawhinTalebi]

Great question. The ICA implementation here introduces randomness into the code. Quickly skimming the MATLAB doc, it's not immediately obvious how one can control it, but I have provided the link below for you reference. www.mathworks.com/help/stats/rica.html#bvnjjp8-5 Hope that helps!

[@manden_fruta]

Thank you! @@ShawhinTalebi

[@RahulSingh-up8jo]

Thank you!

[@ShawhinTalebi]

Happy to help!

[@manjukh22]

How to apply PCA and ICA for cognitive radio? Spectrum sensing and channel detection.

[@ShawhinTalebi]

Hi Manjunatha, Thanks for the question! I'm sorry, I've never worked with that kind of data before. However if you have a set of time series like with EEG, the code I've shared here should work. Now, whether that gives anything useful is another matter. Sorry if that wasn't helpful. Perhaps someone with more experience with cognitive radio will chime in. Shaw

[@manjukh22]

@@ShawhinTalebi Thanks for the reply sir. Great explanation. I will follow your instructions in the reply.

[@shruthivenugopal5792]

Is there any package to implement ICA in skicit -learn ?

[@ShawhinTalebi]

Hi Sruthi, thank you for the comment. Yes, sklearn has an ica implementation. Here is a link to the doc: scikit-learn.org/stable/modules/generated/sklearn.decomposition.FastICA.html

[@me-ou8rf]

If I have EOG channels in my EEG data which already contains Eye Artefacts better then should I remove those channels before applying ICA ?

[@ShawhinTalebi]

Good question. I think you can leave those in. It may even improve the results. Alternatively, you can leave it out and use it to pick out the ICs which contain blink artifacts.

[@me-ou8rf]

@ShawhinTalebi thanks for reply. Is using ICA really necessary ? In many MI-EEG paper, they don't seem to use ICA for their deep learning architecture. They all use BCIC-IV 2a dataset. Is it because during Motor imagery task, which involves short time like 4s , not much eye blinking going on ? Or is the convolution layer automatically take care of that ?

[@silvyarahmatiaraputri5133]

hi, can ICA be used on NASA IMS bearing dataset?

[@ShawhinTalebi]

Hi Silvya, thanks for the question. I'm not too familiar with that dataset. However, if it is a multivariate time series dataset, ICA is typically applicable.

[@ardan7779]

Thanks!

[@ShawhinTalebi]

Thanks for watching 😁

[@JoshBugg]

When you drop the blink components, are you dropping the columns or zeroing them out, as it says in the commented code?

[@ShawhinTalebi]

Thanks for the question. I zero them out to keep the matrix the same shape. Otherwise, there might be issues in reconstructing the signal.

[@JoshBugg]

@@ShawhinTalebi Thanks for clarifying!

[@danielkrupah]

Please can you help me to estimate PCA for panel data using Stata. Thank you

[@ShawhinTalebi]

Hi Daniel, I’m sorry I do not have experience with Stata. However, the PCA video may be helpful: kzbin.info/www/bejne/jXXNq5qkpt6tatU Hope that helps!

[@sfnembedded]

You have explained it so well, thank you. May I ask? Removing blink or the other noise, Can we do it with some combination such as HPF (High Pass Filter), LPF (Low Pass Filter), and BPF (Band Pass Filter) or there's any special use for ICA or PCA? in case of filtering signal. Thank you!

[@ShawhinTalebi]

Absolutely! PCA and ICA are just another set of tools for the signal processing toolbox. So they can be combined with other filtering approaches. For example, one idea is to take a signal, copy it multiple times, and apply a different filter to each copy. Now you’re set up to do a PCA or ICA or both on this set of signals. Hope that answers your question. Please let me know if I missed it.

[@sfnembedded]

@@ShawhinTalebi It's clear right now. Appreciate it! Can you make a video about the theoretical way? I'm very curious about the signal analysis method

[@ShawhinTalebi]

@@sfnembedded Thanks for the suggestion! I've been considering making an appendix to this series, and that sounds like a good fit for it.

[@sfnembedded]

@@ShawhinTalebi Thank you. Already subscribed, can't wait for more great content

[@ShawhinTalebi]

​@@sfnembedded Thank you, I appreciate it!

[@malcolmudeozor3264]

Hey! Great video! it seeams like your data is still autoscaled after transforming it back to its original form. Would you recommend un-Zscoring the data?

[@ShawhinTalebi]

If you are going to use the raw values directly then yes, but if you are going to convert into spectral bands (eg alpha, beta, etc.) then I don’t think that’s necessary.

[@alihamid8085]

Could you please send me a link for EEG dataset description that used in this example.

[@ShawhinTalebi]

Example code and data can be found at the GitHub linked in the description. Data is from me sitting quietly with my eyes open. Hope that helps!

[@tanvikapdi1422]

can u pls guide how eeg can help with data related to brain disfunction

[@ShawhinTalebi]

Thanks for the comment! I’m sorry that is not my area of expertise. I’m sure there is a good literature focused on that.

[@bluehope4143]

Can you apply ICA for sharp spike noise removal from local field potentials?

[@ShawhinTalebi]

That's a good question. The blink artifacts here are sharp spikes, and ICA does a decent job at handling them. But I'm sure there are cases where this breaks down. It's definitely worth a try!

[@bluehope4143]

@Shawhin Talebi I gave it a try, it removes many spikes but not all. It was still very helpful and reduces my preprocessing work. Thanks!

[@ShawhinTalebi]

@@bluehope4143 Good to know, thanks for reporting back :)

[@PRASANTHPR-v1g]

❤❤❤❤

[@KongQuestCo]

I was gonna ask if they could use the eeg with their eyes closed. Then I remembered how you were getting the data lol

[@ShawhinTalebi]

Yes

[@עינת-ל]

Thanx

[@dtr_cpg]

Cool

[@pravinandhale]

❤

[@kumarpratap5115]

so ICA is used to separate the sources? and remove noise?

[@ShawhinTalebi]

Yes that's the idea. We separate the signal into constituent sources, some of which we can pick out as noise.

[@physicswithamartya1061]

Sir can you provide a pdf note regarding it...

[@ShawhinTalebi]

Hi Amartya, thank you for the comment. The blog dives a bit more into the details. Hope that is helpful! Blog: towardsdatascience.com/independent-component-analysis-ica-a3eba0ccec35

[@physicswithamartya1061]

@@ShawhinTalebi 😇

[@arvoli2596]

Wow

[@ShawhinTalebi]

🙌

[@ITEMAE]

分り易い！

[@ShawhinTalebi]

Glad to hear, thanks for watching!

[@fa7234]

This is very similar to wavelet noise reduction

[@GregoryMonk-o5b]

Ralph Place

[@ShawhinTalebi]

📰Read more: towardsdatascience.com/independent-component-analysis-ica-a3eba0ccec35?sk=6ddb9ae6d343de7f667255ab564099dd