Science topic

ICA - Science topic

Explore the latest questions and answers in ICA, and find ICA experts.
Questions related to ICA
  • asked a question related to ICA
Question
4 answers
In my experiment, I have two groups of subjects (20 in total, 10 in each group) and I want to compare the activity of DMN between two groups. And for each subject, I have four BOLD fMRI scans.
My question is that should I input all of them (80 scans) into Melodic and average these repeated scans after doing dual regression or for each subject combine 4 scans into one scan with a long timeseries before doing ICA?
Any suggestion will be appreciated.
Relevant answer
Answer
Sorry. Since for now I do not analyze the fMRI imaging data any more, I cannot give you any good suggestions about the analysis. BTW, thank you for your interset. Nair Ul Islam
  • asked a question related to ICA
Question
2 answers
Hi all,
I have a MEG data set which is about 5 minutes long for each subject. I want to detect and correct existing artifacts using the ICA approach using the Brainstorm toolbox. When I performed ICA, the first component (the dominant one) has a very small amplitude, (about 1000 times smaller in comparison to other components) and looks like a DC signal. Should I remove this component? It is the first component and I want to be sure about this decision as it may be a signal, not an artifact. In the following, you can see the components.Thanks!
Relevant answer
Answer
You should not eliminate unless you can identify a source for the component. See for example https://www.ee.columbia.edu/~dpwe/papers/HyvO99-icatut.pdf, Section 7.1, page 22 in which the authors discuss an MEG experiment in which the test subject was asked to blink and make horizontal saccades, then to bite teeth for as long as 20 seconds. In one version, a digital wristwatch was placed in the Faraday cage, and it produced an artifact.
Question: Is the signal of low amplitude very regular, even more that the ECG signal of this paper? If so that is strong evidence that it is an artifact.
  • asked a question related to ICA
Question
2 answers
Hi,
I'm preprocessing some EEG datasets in an EEG-VR experiment with some neck movements.
In some of the datasets, some spindle-like alpha activities appear in most channels (mostly occipital, parietal, and frontal but less central channels). I'm not sure if they're some kind of noise maybe because of the VR headset on the EEG cap and therefore should be removed or not.
After doing ICA, these activities will also appear on some of the components, which ICLable identifies as either 100% or above 95% brain activity. I attached some photos of the channel activities and components. I'll appreciate it if you can help to be sure whether these are noise or not and if they are how should I remove them.
Thank you very much.
Relevant answer
Answer
Alpha appears often when participants get tired and sleepy, even with eyes opened. It *is* brain activity, but depending on what you want to do with your signal, you may want to filter it out. E.g., in ERP research we normally get rid of segments with more alphas, as they affect strongly the averaged waves. On the other hand, other researchers may be interested in alpha specifically.
So, shortly - it is not noise in the sense of external signal or artifacts, but depending on what you want to do with your EEG data, you may want to remove it.
  • asked a question related to ICA
Question
2 answers
Hi everyone, I'm assessing one's cognitive performance based on their brain activity recorded through an electroencephalogram (Emotiv EpocX EEG headset). Recently there is a very high incidence of bad EEG quality during the recording and about 70-80% of dirty data have to be removed. The test was about 120s long and left with <30s after dirty signal removal. May I know how do I determine the minimum EEG data points in that particular test to perform ICA? And how do I confirm if the remaining data is representative due to very limited ''clean'' data after noise removal?
Thank you.
Relevant answer
Answer
Hi Cheryl,
the number of time points needed for sufficient ICA descomposition can be estimated with the numbers o eletrodes, please see Onton et al., 2006 ( ) for details.
  • asked a question related to ICA
Question
1 answer
Although I went through the literature, I couldn't understand the identification of the bifurcation point between ECA, ICA and the methodology of filament insertion. It will be helpful for me if you can share a video of the procedure
Relevant answer
Answer
Hi Lingesh Allakonda, hope you will find this txt and video useful,
also, pls have a look at this article
best wishes
  • asked a question related to ICA
Question
3 answers
Hello,
I have 19 EEG channels and use ICA for EEG artifact removal.
Which criteria can be used to remove or accept EEG channels using EEGLAB?
Thanks
Relevant answer
Answer
Dear Pedro Morais,
Thanks for your reply
Most of artefact removal techniques use visually interpretation. I am looking for statistical criteria
Kind regards
  • asked a question related to ICA
Question
3 answers
Hi, I am currently working on a project which requires me to implement the ICA algorithm in real time, to be specific, I am trying to separate the noise from the audio. When I perform the algorithm in offline, it works fine despite the amplitude of the separated audio is a bit soft. However, when I implement it in real-time, the separated audio becomes very soft. Any source code that I could refer to solve this kind of problem. Thanks.
Relevant answer
Answer
What you are describing appear to be two different problems, one compounding the other
For the first problem
1) Which ICA implementation are you using?
For the second problem
1) For the real time capture what sampling rate are you doing the capture?
2) what are you using to capture the signal in real time(e.g. are you using ALSA on Linux or are you using Windows)?
Note that the RM Cortex-M7 of the Teensy 4.0 is a slow processor compared to an Intel or AMD
  • asked a question related to ICA
Question
3 answers
Hi!
I am currently doing a intercoder reliability test for my phd. I have conducted 17 semi-structured interviews, and segmented/coded them in atlas.ti.
Each code has a upper category it belongs to, for example subcodes “parent”, “uncle”, “child” belong to the upper category “family”. I have allowed sub-codes within the same upper category to overlap while coding, tagging for example both “parent” and “child” in instances where they are both mentioned in the same segment.
I have had a second coder (me being the first) code a sample of the data for a ICR test.
After coding, I realized Cohen’s Kappa and Krippendorff’s Alpha as methods do not allow the usage of more than one code per category for each coded segment. I should have checked this prior to coding, and i now have to work with what i have.
So, I am now asking for advice on which method, or methods to apply in this situation? I am opting to use two, but I don’t know which one’s to use.
I could use a “Fuzzy Kappa” model, which is modelled after Cohen’s Kappa to suit these kinds of overlapping cases. There does not appear to be a standard interpretation of Fuzzy Kappa however. In addition, Cohen’s Kappa in itself still faces some criticism for its shortcomings, so im not completetly sure about this. Another option is for me to calculate a basic cohen’s Kappa on each code separately and then take the average of that, this is an advice I have seen on this forum earlier.
I could also go back to my data and split the codes/ redefine the upper categories to avoid overlap so I could attempt a Krippendorff’s Alhpa test. However, I am not sure if I could get the second coder to come back for splitting their codes, and redefining the categories would be counterintuitive. Krippendorff’s Alpha seems like a more reliable method, but it also has stricter requirements than my resources may take ( requires specific sample sizes, does not allow for the primary researcher/code developer to be one of 2 raters).
Any ideas of which methods to use/ combine are of great help!
Relevant answer
Answer
Yes, you can take the average of the separate reliabilities. I would only use Cohen's Kappa for codes that appear in between 30 to 70% of the segments, because otherwise it has severe problems due to "proportional splits" on the margins. I don't know if Krippendorf's alpha resolves that problem or not.
  • asked a question related to ICA
Question
4 answers
I would highly appreciate if you can comment down below some leading works on ICA in the context of cascading disasters (disasters whose effects spread through space and time).
I prefer numerical modeling works.
Many thanks!
  • asked a question related to ICA
Question
2 answers
I have simulated fmri image. Also performed the ica using GIFT tool over the simulated image. I would like to know how to perform comparison between these two to know the performance level of the ica algorithm used.
Relevant answer
Answer
Hello Mr. James
Thank you very much for your suggestion.. Definitely I will try that..
Regards
Purnima
  • asked a question related to ICA
Question
1 answer
Hello! I have rsfMRI data from one subject with two conditions.
I need: 1) to identify major networks 2) compare these networks between two conditions.
Do you think ICA is appropriate method to use? If yes, should I run ICA for each condition separately and should I restrict the number of components?
I tried to run ICA for one of the condition, and I got results with 200 components, looks like a problem with overfitting (attached). This is my first time when I work on ICA, and I do not know how to interpret it.
Relevant answer
Answer
This would be a great question to post in our new free medical imaging question and answer forum ( www.imagingQA.com ), there are already a few fMRI questions on there. If useful, please feel free to open a new topic at the link below :
  • asked a question related to ICA
Question
5 answers
what is better: duplex for CCA or ICA or ECA?
Relevant answer
I agree with Benito jonas Baldauf.
To add, the measurements should be taken at the anterior, posterior, medial and lateral walls and the average of the measurements to be taken.
  • asked a question related to ICA
Question
4 answers
I have data from 6 EEG channels (F3-A2, C3-A2, A1-A2, F4-A1, C4-A1, A2-A1) sleep data. And 1 channel for ECG and EOG, are these enough to run ICA to remove artifacts caused by EOG and ECG from the EEG channels?
Also which algorithm would be recommended to use?
Relevant answer
Answer
Uttara u Khatri : Yes that should be sufficient. But when PSG signals are considered for analysis. It also depends upon how many features you have extracted from each of these channels. Make sure all the features of PSG contribute equally in the analysis.
  • asked a question related to ICA
Question
2 answers
Greetings RG community,
I have been working on a pipeline for the classification of EEG motor imagery signals. This is currently being done on the Giga Science MI dataset with 52 subjects, using all 64 EEG channels. The question can be isolated to my first preprocessing step involving ICA by way of Hyvarinen's fast fixed-point algorithm. If I develop spatial filters (vectors of the unmixing matrix) using only the intended training data, is it a violation of appropriate protocol if I then project all raw data (which includes testing data) on these vectors in an attempt at blind source separation? The nuanced thing that brings about concerns is that the raw data is provided in two matrices each containing LH+RS and RH+RS signals (LH = left hand; RH = right hand; and RS = resting state). If the spatial filters wL and wR were constructed using the LH and RH training data respectively, then the original raw data of the LH and RH matrices (including both training and testing data) were projected into these directions prior to the partitioning of trials, is this considered using knowledge of the classes thereby rendering the entire analysis ineffective? At first, I thought I was in the clear because class labels were not used as ICA is unsupervised, now I think it pertinent to ask someone that may have experience in this field.
My results were great under these conditions (perhaps this would be obvious). To check if I could replicate results a different way, I vertically concatenated the LH and RH training data (doubling the number of samples in comparison to the conditions described above) and developed a single matrix of spatial filters then projected all original data onto these but the results were poor, indicating a large drop in spatial resolution. Ideally, I would like to develop a single set of spatial filters that can be applied to all data indiscriminately, if anyone has any advice given the situation it would be greatly appreciated. Since this step is being done prior to the partitioning of MI trials, I have considered performing ICA on some vertically concatenated trials after partitioning and was wondering if this would yield good results as I have also read that the resting state signals contain important information for the minimization of mutual information (maximization of differential entropy), so I am uncertain with this approach. I am also in the process of replacing FastICA with SOBI, JADE, and infomax in an attempt to gain higher spatial resolution. Please excuse any off-putting terminology as I recently pivoted from functional protein dynamics recognition and prediction to BCI-EEG motor imagery classification. Feel free to share any thoughts, all advice is welcomed.
Thank you for your time,
Tyler J Grear
Relevant answer
Lutsenko E. V. Diagnostics and forecasting of professional and creative abilities by the method of ACK-analysis of electroencephalograms in the "Eidos" system / E. V. Lutsenko, A. N. Lebedev / / Polythematic network electronic scientific journal of the Kuban State Agrarian University (Scientific Journal of KubGAU) [Electronic resource]. - Krasnodar: KubGAU, 2003. – №01(001). P. 59-61 – - IDA [article ID]: 0010301009. - Access mode: http://ej.kubagro.ru/2003/01/pdf/09.pdf, 0.188 cu. p. l.
EEG forecast the success of psychomotor test while reducing the level of wakefulness: an analysis of the results of the study / T. N. Shchukin, V. B. Dorokhov, A. N. Lebedev, E. V. Lutsenko // Polythematic network electronic scientific journal of Kuban state agrarian University (Scientific journal of Kubsau) [Electronic resource]. - Krasnodar: KubGAU, 2004. – №04(006). P. 290-306. - IDA [article ID]: 0060404022 – - Access mode: http://ej.kubagro.ru/2004/04/pdf/22.pdf, 1,062 cu. p. l.
EEG forecast the success of psychomotor test while reducing the level of wakefulness: a description of the experiment / T. N. Shchukin, V. B. Dorokhov, A. N. Lebedev, E. V. Lutsenko // Polythematic network electronic scientific journal of Kuban state agrarian University (Scientific journal of Kubsau) [Electronic resource]. - Krasnodar: KubGAU, 2004. – №04(006). P. 277-289. - IDA [article ID]: 0060404021 – - Access mode: http://ej.kubagro.ru/2004/04/pdf/21.pdf, 0.812 cu. p. l.
EEG forecast the success of psychomotor test while reducing the level of wakefulness: statement of the problem / T. N. Shchukin, V. B. Dorokhov, A. N. Lebedev, E. V. Lutsenko // Polythematic network electronic scientific journal of Kuban state agrarian University (Scientific journal of Kubsau) [Electronic resource]. - Krasnodar: KubGAU, 2004. – №04(006). P. 268-276. - IDA [article ID]: 0060404020 – - Access mode: http://ej.kubagro.ru/2004/04/pdf/20.pdf, 0.562 cu. p. l.
  • asked a question related to ICA
Question
2 answers
I have used ICA for the problem of mixing acoustic and UHF signals from partial discharges. I know that it has been used in Biomedical, Acoustic, and many other fields. Do you comment on how you have used it?
Relevant answer
Answer
I have used it for feature selection purposes in my study but did not work better than other methods (PCA and LDA).
rica(X,q,Name,Value is the implemented function in MATLAB2019.
  • asked a question related to ICA
Question
3 answers
Hello, I used EEGLAB interface to run ICA and remove artifact components, after which I saved the dataset in a .mat format. When I load the set, the structure contains the file name under 'data' in .fdt format. How can I access that file? I have tried loading it separately as an ASCII, but it results in an error: 'Unknown text on line number 1 of ASCII file p1_ICA.fdt' Is there any other way to access the cleaned , post ICA, data? Thank you, Rawan
Relevant answer
Answer
Thank you so much
  • asked a question related to ICA
Question
12 answers
I need MATLAB, C# or R code of above mentioned models. How could I find those?
Relevant answer
  • asked a question related to ICA
Question
3 answers
Dear all,
I recently tried to optimize my scripts to preprocess EEG data using EEGLAB2019. I found that including EOG channels in ICA significantly improved the subsequent ERP results, i.e., the classic LPP appeared. However, if I included EOG channels in ICA, the subsequent average re-reference had to also include them. If I performed average re-reference on only brain channels, the eeg_checkset would report an error ( ICA Index exceeds matrix dimensions). If I excluded EOG channels in ICA, average re-reference on only brain channels worked well, which, however, generated a bad ERP wave.
The figure attached is the ERP waves based on four-subject preprocessed data. The only difference between them is whether the EOG channels were included in ICA. ICALabel was used to automatically remove the EOG components.
Relevant answer
Answer
its generally not ok to include EOG EMG channels in the rereference, otherwise it would generate huge artifacts in your EEG. On the other hand its beneficial to include when you run the ICA since the the high amplitude of eye-artefacts can help get rid of the artefact-related components in the EEG
  • asked a question related to ICA
Question
2 answers
Good morning!
We are planning to run an fMRI experiment on patient. I would like to add a resting state acquisition. However, to reduce to the minimum the time in the machine for the patient, I was wondering the minimal number of acquisition necessary if I am planning to analyze the data with an ICA. Moreover, I know that my group of patient will be not really big (probably around 20 so the double with my control group).
Thanks a lot your insight and feedback!
Relevant answer
Answer
Bonjour Isabelle!
Thanks a lot for your answer. 10mins is what I had in mind.
I was wondering because I read two articles with acquisition of 4 and 5mins and to me, the number of time points was too low.
Again thanks for your help and advises!
  • asked a question related to ICA
Question
1 answer
I have been working with FSL software (melodic & dual_regression) to parcellate intrinsic connectivity networks both at a group and subject level. I successfully ran dual_regression and have the outputted z-maps for each subject and component. I want to create a summary statistic for each outputted z-map so that I can compare differences across participants using a dimensional framework. I am not interested in analyzing subgroups, even though all the FSL documentation is all about group differences.
I am wondering how I could calculate a summary statistic for each z-map that represents the network strengthen/integrity of a given subject. I basically want to know how over-expressed or under-expressed is a given network is for each subject. I want to use this approach to test a hypothesis regarding individual differences in symptoms of depression. I read this in a few books and papers, but no one has explained how to quantify these subject-level networks from nifti files into numeric summary scores that could be inputted into a multiple regression.
Attached here is my group ICA output that was manually capped at 30 components (melodic_IC.nii.gz) and a couple dual regression output files (original and z-maps) from two subjects to give you an idea about what I am working with. The 3rd component is of most interest as it most resembles the default mode network. Please help someone!
Relevant answer
Answer
Upon more thinking and reading, I think the easiest solution is to use fslmeants to simply extract the mean beta weight across all voxels from the Stage-2 Output Maps. I did multiple sanity checks and find that the mean values and individual beta maps are fairly comparable (see bottom row of attached screenshot).
It also helps to apply smoothing prior to group ica (kernel = 3mm) to get cleaner components that more resemble functional networks in case anyone runs into this problem in the future.
  • asked a question related to ICA
Question
8 answers
Have implement neural network based mask detection algorithm with different colors of masks and that is working with 99% accuracy (around 2 fails in 850 tests) but the CNN based algorithm is too slow on boards like Raspberry Pi around 1-2 FPS at 720p. Shall I try PCA /ICA based techniques?
  • asked a question related to ICA
Question
6 answers
We have tried various modifications to our standard Western Blotting procedures now trying to detect CHOP (GADD153) in our human cell lysates after e.g. stimulation with 1 µg/l tunicamycine, but to no avail.
Do you have some helpful pointers as to how to detect CHOP?
We have used nitrocellulose and PVDF membranes in the past. Primary antibody is the recommended Santa Cruz monoclonal mouse anti-GADD153 (B-3) at 1:250. For technical reasons, we don't do the transfer on ice, however, actin signal is always strong on our membranes.
Is it a matter of amount of protein loaded? We have done roughly 20-30 µg/lane so far.
Thanks for any good ideas or working protocols!
Relevant answer
Answer
I stumbled upon this conversation as we were also experiencing difficulties in detecting CHOP protein. For your information, using a fixation step with glutaraldehyde optimized the detection of CHOP on Western Blots.
cf: "Improvement of immunodetection of the transcription factor C/EBP homologous protein by western blot"
  • May 2020Analytical Biochemistry
  • DOI: 10.1016/j.ab.2020.113775
  • asked a question related to ICA
Question
2 answers
What is mean by Eigenvalues in ENVI, when performing PCA or ICA operations in ENVI environment? And what is mean by Covariance Matrix or Correlation Matrix statistics?
  • asked a question related to ICA
Question
5 answers
Hi folks,
I am using 14 channels of EEG system.
In the step of artifacts removal, I wonder how to detect it for low resolution (14 electrodes) system. I’ve tried ICA, but it doesn’t give good result.
Would appreciate if you could attach any published papers/conference papers that I can follow.
Would appreciate any ideas/suggestions with it.
Thank you so much!
Relevant answer
Answer
ICA is not good with few channels. Thresholding based on extreme values will do the job. A rejection based on visual inspection is also in most cases a good solution.
There are some good tutorials on how to do this with the various toolboxes for EEG analysis:
  • asked a question related to ICA
Question
3 answers
I am working on optimizing well placement in the condensate reservoir model using an algorithm. Any kind of code example will be appreciated.
Relevant answer
  • asked a question related to ICA
Question
11 answers
hi
I am working on EEG's data and I am confused about some of the components. I know the eye blinking one and some others but still I have a doubt on some of them. I use adjust program but I can't trust it due to its some mistakes that I have been seen from it.
I would be so happy to help me and tell me if any of this ICA components in these 4 pictures should be removed or not? please tell me your reason. (I will put component activation picture of them if you need)
thank you
Relevant answer
Answer
Following on from Luca Bischetti , rather than just using the ICLabel website for training, you can use the ICLabel EEGLAB plug-in, which uses the ICLabel website and machine learning to label your ICs.
It also rather helpfully plots the topography alongside the time series, continuous frequency and spectrogram data.
  • asked a question related to ICA
Question
3 answers
Hi friends, Can you Please tell me analysis of EEG signal using independent component analysis using MATLAB. If possible can you give me the source code for EEG analysis using ICA.
Relevant answer
Answer
Hey
Firstly , what kind of EEG analysis u want to make using ICA
  • asked a question related to ICA
Question
4 answers
Hai friends,
I have started my research recently, working on Artifacts removal EEG signal. I have done the literature survey on PCA, ICA,CCA, EMD,EEMD. I have understood somehow. Can you please tell me how to take EEG signal as input and how to get EEG database.
Relevant answer
Answer
You can try browsing through Brainstorm software website https://neuroimage.usc.edu/brainstorm/Introduction . Another resource you can make use is Mike X Cohen lectures http://mikexcohen.com/lectures.html .
  • asked a question related to ICA
Question
3 answers
Hi all,
I have this doubt that I could not solve by myself. I'm performing time-frequency analysis with EEGlab software for my master thesis and I'm not sure whether it is more correct to perform ICA and epoching in a precise order. I found some reccomendations in EEGLAB tutorial manual saying that it is advisable to perform ICA on a greater amount of data as possible. However, if I run ICA on EEGLAB before epoching, I cannot have a good visualisation of the ICA components (only the maps and components scroll). Instead, I have seen in a tutorial by Mike Cohen, that he performs epoching before ICA, so there is the possiblity to look at each component latency and frequency before rejecting. Personally, I found myself better with the second method, but I am not sure wheather it is correct or not.
I hope you can help me in this dilemma! Thank you!
Relevant answer
Answer
Hello,
Steven Luck, who is pretty widely regarded as the god of ERP analysis methods recommends that people first visually inspect and reject highly abnormal data (like a period of extreme muscle noise), then perform ICA, then epoch.
Here is a post about it on his site with a full outline of his recommended data processing/analysis steps: (The clearest step by step outline is near the bottom of the article)
  • asked a question related to ICA
Question
4 answers
I am using a 24-channel system with 19 active electrodes to record the brain activity during flanker task. I run ICA to remove ocular artifacts. Although ICA can identify 1 or 2 eye-blink components, time course and ERP image show there still task-related brain signal in these eye-blink components. My question, is there any tips that can help to keep the real brain signal?
Relevant answer
Answer
I agree with Franklin Lue. ICA "takes the signal apart" and identifies components (also artifacts embedded in data) that are independent of each other, but of course different parameters of ICA may result in variations in your data. Good news is that blinks and saccades are way stronger than "the real brain signal", so it's easier to identify them. They also influence more the frontal electrodes than the central, parietal or occipital ones, so if you're interested in components in the latter locations, it's not such a big deal. I wouldn't worry about data loss caused by rejecting [artifact] channels with ICA, it's a risk the field takes, but if you're not convinced (and of course there are valid arguments for this scepticism), dig into the literature to understand the algorithm better and make an informed decision about how you want to treat your data. A brief search led me to this article, maybe you'd find it useful:
  • asked a question related to ICA
Question
4 answers
e.g. default mode to performance on attention tasks, salience network to emotional salience tasks, and executive motor component to executive tasks.
Relevant answer
Answer
Thanks but that is exactly the problem. The ICA components are being ascribed functions on indirect evidence and no direct evidence exists.
  • asked a question related to ICA
Question
4 answers
I am working with data that is very messy such that, for a given file, the decomposition will contain a noise artifact IC that accounts for >70% of the variance. Out of curiosity, I reconstituted the data without that high variance noise IC (the reconstituted data still looked very messy), then reran the reconstituted data through ICA and achieved a much more reasonable decomposition and fairly clean reconstituted data. Any thoughts as to whether this is an allowable method for cleaning data?
Relevant answer
Answer
As I said, I don't have a deep mathematical understanding of ICA, but I think the critical point is this: when you start with 64 channels, then your data has a maximum of 64 dimensions, i. e. 64 components will definitely be enough to convey the full information contained in your data. When you remove one (say, noise) component, only 63 dimensions are left. But when you back-project this to electrode space, you pretend that your data is 64-dimensional again, bc you want to restore your 64 channels. I think the mathematical expression would be that your data matrix is rank-deficient now (not sure if it's the correct term though). So you extract 64 components again, while there are really only 63 dimensions left in your data.
Anyway, I think that you can get a more mathematically substantiated answer if you repost this to the EEGLAB forum or FieldTrip mailing list, as Georgia suggests.
Cheers,
Marius
  • asked a question related to ICA
Question
4 answers
I usually read something along the lines of "the fewer the better".
But what if you have multiple components that seem to contain artifacts?
Relevant answer
Answer
I agree with @Mathew B Wall
  • asked a question related to ICA
Question
3 answers
I have a dataset of an EEG recording referring to a task characterized by several "rest periods", during which the recording is not interrupted. For this reason, I was thinking that removing all these periods by epoching the recording and time-locking it to my ERPs of interest would facilitate the ICA by isolating only the artifacts occurring during my time-window of interest. However, I was warned to be careful with this operation since it can cause artifacts occurring along different epochs to overlap with each other. Any suggestion about the feasibility of this approach would be greatly appreciated. Thanks
Relevant answer
Answer
You can run ICA before or after epoching. But ICA is known to work best with more data. Your continuous raw data gives more information for the ICA to work on to get you more stable components (ICs). Your epoched data has already lost some information due to pre-processing (windowing etc). So, ICA can only work on so much. Always remember “garbage in garbage out” principle.
  • asked a question related to ICA
Question
4 answers
I did FSL FEAT on resting fMRI with TR=2s, 3x3x4 mm^3 resolution. What is the best way to remove noise ICA components? I tried ICA-FIX (using standard trained data) and ICA-AROMA. the results are quite different. FIX identified only couple noise components while AROMA detected a lot noise components. Any suggestion which one to choose and what is the best way to use FIX and AROMA? any other alternative beside manual removal? Thanks.
Relevant answer
Answer
You bet. I tried both AROMA and FIX With the trained weights from package. Fix picked up almost all signals and some noises while AROMA missed a few signals. Therefore, I will go FIX for my scans now. thanks all.
  • asked a question related to ICA
Question
1 answer
Hi,
is there an existing Python implementation of Fourier ICA?
I only found the original MatLab implementation of Hyvärinen at
Thanks!
Relevant answer
Answer
Dear Armin Goudarzi,
Entire code is there in MATLAB then why you can convert it for python!!.
  • asked a question related to ICA
Question
4 answers
I've collected EEG data using 24 channels system. Twenty-three channels (19 active EEG, 1 VEOG, 1HEOG, 1 right-earlobe) were referenced by left-earlobe, and one channel used for EOG bipolar (V-EOG referenced by H-EOG). I'm trying to clean my data using ICA.
My question, what is the minimum number of EEG channels that required to properly run ICA?
and, is it better to run ICA with all the channels (24) or i need to remove the bipolar EOG channel that not sharing the same common reference?
Thank you,
Relevant answer
Answer
Could you say a bit more about what you are trying to accomplish with your ICA? Some people use ICA as an artifact removal technique to extract muscle or blink artifacts from their scalp channels, others will use ICA to extract task- or experiment-related brain data like ERP components, etc.
From what I know, there isn't really a minimum number of channels necessary to run an ICA, but you will only get back as many ICA components as you have channels. So, if you entered all 24 of your channels, you have the potential to examine 24 ICA components. The more channels you have, the more you can "dissect" your data into unique aspects.
If you are looking at the brain data, I would recommend NOT including the EOG channel. ICA is dependent on commonalities in your data, and if the EOG is not referenced the same way as your scalp sensors then it will likely throw off your results and introduce measurement error.
  • asked a question related to ICA
Question
3 answers
Hi everybody,
I am using a 32 electrodes EEG from Brain Products GmbH. The electrodes of this device are all connected to a splitterbox, that in turn is connected to the amplifier. I have not the adapter required for the EOG electrodes to be connected to the amplifier, then I will have to use the cap's electrodes as EOG.
Can I use these electrodes as EOG? The reference for the EOG will then be the same as the one for the electrodes on the scalp? How many of these electrodes should I use as EOG? Can I use this kind of EOG for running an ICA for eye artefacts removal? Or can ICA be also run without EOG?
Thank you very much,
Alessandro
Relevant answer
Answer
ICA is an unsupervised method witch means noisy signal is its input and ICA out put is cleaned signal and noisy. So it is Not essential to extract artifact individually from other channels (EOGs channels for example).
what I am trying to say is that, practically EOG artifacts just appears in frontal electrodes Not hole 32 electrodes. However I have a suggestion pre-processing strategy for you in following. First, pas all of 32 signals trough high-pass frequency filter (6th order, butterworth, 0.5 Hz cut-off). After that, applying ICA to hole them all. In addition, if there are moving artifacts in your recording protocol (your subjects are not in resting position), you can consider a 200uV threshold on the EEG amplitude signals before applying ICA.
It is a very common approach in noise and artifact cancellation for EEG signal. you can find it in pre-processing stage of many authentic papers.
  • asked a question related to ICA
Question
3 answers
Hi,
I am trying to find templates/mask of the DMN, saliency network and attention network.
I did an ICA analysis and want to be able to classified my different ICA in specific networks...and eventually quantify the overlap.
Does this kind of template/mask exist? or do I have to create them myself?
Thanks a lot for your help!
  • asked a question related to ICA
Question
6 answers
Hi,
I have a question about EEG signal processing in EEGLAB. I am not sure whether the preprocessing steps that I did are suitable. It would be great if someone can look at it and maybe give me tips on how I could improve my signal preprocessing.
I did a relaxation task and a attention test using the Emotiv EPOC+ device. Now, I want to find out if the ratio of the band power alpha/theta are different in these two phases. For that reason I processed the raw EEG signal as followed:
1. Import raw data
2. read channel locations
3. FIR filter: High-pass filter at 0.16 Hz to remove background signal and DC offset, Notch filter at 50 Hz to remove the interference
4. Scrolling the data and rejecting clearly bad stretches but ignore the eye blinkings
5. ICA
6. Reject components
7. Scroll the data and remove signals which exceeds 100 microvolts interval
8. FFT
9. Filtering the alpha and theta band
Thanks for your help
Relevant answer
Answer
Hi Tanja,
based on my modest experience with ERPs using Brainstorm (but shouldn't be much different from EEGLAB), I think your preprocessing pipeline makes sense.
If you haven't done that already, I can suggest you check out some EEG videos from Mike Cohen's channel (https://www.youtube.com/channel/UCUR_LsXk7IYyueSnXcNextQ), and this one on pre-processing and cleaning in particular: https://youtu.be/uWB5tjhataY
The steps I learned to follow from my awesome EEG mentor (he's got some cool EEGLAB and R resources you may also find useful here https://sites.google.com/site/giorgioarcara/erpr) are:
1) manually (or with code) remove the messy parts of the signal at the beginning and at the end (this later leaves more room to the ICA to find the less detectable artifacts)
2) apply high-pass @ 0.5 Hz
3) run PSD and look at the raw signal to identify bad electrodes and remove them
4) apply low pass @ 40 Hz (so no need to notch-filter)
5) ICA + reject components
I usually stop here with the pre-processing of the raw signal before epoching for ERP extraction and reject single bad trials, so up to here you would be good to go with your step 7.
Hope this helps.
Cheers,
Tiziano
  • asked a question related to ICA
Question
5 answers
Hello,
I have a question regarding synchronization of artifact rejection in eeglab. I use the version: "eeglab 14_1_1b" (MATLAB plugin)
I work with epoched data. To find artifacts like blinks and eye movements/saccades I have used the built-in ICA function and followed the recommended guidelines from Chaumon et al. 2015 to pick out components. Furthermore I have used the "simple voltage treshold" til sort out any voltage below -75 microvolt and voltage above 75 microvolt.
My issue arises when I want to synchronize the artifact info in EEG and EVENTLIST. This is a must do step in order for me to compute the average ERPs. eeglab won't let me do this, and it brings me this exact message when i try to synchronize:
"It looks like you have deleted some epochs from your dataset. At the current version of ERPLAB, artifact info synchronization cannot be performed in this case. So, artifact info synchronization will be skipped, and the corrosponding averaged ERP will rely on the information at EEG.event only. Do you want to continue anyway
(yes, no)." I need to find a solution, so that every rejected trails and removed components will be removed from the dataset when averaging the ERPs.
I have not been able to find any version of eeglab which seems to solve the problem.
Any suggestions?
Best regards
Morten
Relevant answer
Answer
You're welcome Morten. Also, it seems from what you are saying that you segmented your data into epochs before you ran the ICA (and thus deleted the epochs containing bad data from your dataset before running the ICA).
My personal advice to you would be to clean your data and run the ICA, then segment it into epochs after you've removed the bad ICA components. This ensures that you get enough blinks/other artifacts so that the ICA can train on them, and increases the chance that you'll get a good ICA decomposition (this method is also recommended by the creators of EEGLAB). Using this method will also avoid the ERP/EEGLAB synchronization issues that you are having, as if you add a binlist after removing data then ERPLAB will only read the events left in the file.
  • asked a question related to ICA
Question
4 answers
I want to know how to extract the most relevant variables of a model using ICA.
When using PCA I can just verify the eigenvalues and select the new variables based on its variance. If I have 100 variables, but only 60 have relevant variances I can ignore the other 40. How does this procedure work in ICA ?
Relevant answer
Answer
  • asked a question related to ICA
Question
4 answers
Dear all,
It seems my data are suffering from alpha distortion. It is possible to reduct it from some data with the help of ICA but for half of the participant ICA does not work that well.
So for now I tried to reject more components than usual (15-18 out of 64 ICs) and run the ICA again if it can catch the alpha. But it doesn't work.
In a nutshell, I wonder what are the best ways to reduce the alpha disturtion from data
Note: I also heart that PARAFAC seems to work better then ICA for reduction of Alpha but I cannot find a way to apply PARAFAC with EEGLAB GUI
Best,
Behcet
Relevant answer
Answer
I do not think it is a good idea to purposely delete ICs from your data that you believe are present due to brain activity. In most cases it is not a defensible assumption that this activity is 100% not relevant to what you are measuring, especially if it is consistent enough to occur during an ERP average. Also, it is not a good idea to re-run an ICA after you have removed ICs because you have reduced the number of dimensions in the data. In this case, you would have to run a PCA and enter the number of components that were left in the original dataset to obtain a valid decomposition.
I am not sure of your experimental design, but one of the best ways to reduce the amount of alpha activity in an ERP average is to use a baseline correction that is a multiple of 100ms. Because the dominant frequency of alpha oscillations is 10Hz, this value will ensure that there are an equal number of alpha oscillations in the baseline which will normally remove most of the alpha waves. If this does not work, you may be able to isolate the alpha waves in an ERP average by using Joe Dien's PCA toolkit to run a temporo-spatial PCA, which you can read your final EEGLAB/ERPLAB files into.
  • asked a question related to ICA
Question
5 answers
Hello everyone,
I wonder how many ICA component do you generally eleminate aproximately? Do you think eleminating 25 out of 64 component sounds reasonable.
Relevant answer
Answer
Dear Behçet,
I think giving general recommendations for how many components to reject is not possible. This is foremost due to differing levels of noise –if noise is very low and stereotypical, few components might capture it sufficiently. If noise levels are high and / or if artifactual sources are diverse, most components likely include some noise.
Reducing noise before ICA is most crucial. As Marius already mentioned, data should be carefully inspected in order to remove non-stereotypical and large amplitude noise. If the gamma band is of interest, separating data into a low (e.g. <30 Hz) and a high band (e.g. > 30 Hz) can also be helpful.
In my opinion, however, one decision has to be made a priori. That is, how conservative or liberal do I want to handle the decision for rejection? The conservative perspective would be to reject components only if there is a very clear signature of noise (e.g. heart beat, eye movement etc.). After all, every component represents a mixture of signal and noise. The liberal perspective would keep components only if noise is apparently absent. While both approaches are legitimate, it is very important to be consistent across participants.
Cheers,
Jonas
  • asked a question related to ICA
Question
3 answers
ICA(Independent Component Analysis) is a computational method for separating a multivariate signal into additive subcomponents that they are statistically independent from each other. If so, it is necessary that there are no similar ICs between each other, because there're independent from each other. But I saw very similar ICs in my part of EEG processing. How do we interpret it?
* I attached a example for better understanding. In pictures, these look very similar.
Relevant answer
Answer
Actually, you should be very cautious about this scenario. The ICs must look like different, however, since, ICA is an approximation, sometimes you will recover similar ICs. Try to increase the number of iterations, or reduce the dimensionality of ICs. Otherwise, similar ICs suggest that with those inputs, there are still inseparable sources. Adding more inputs probably help too.
Good luck!
  • asked a question related to ICA
Question
1 answer
A. I'm trying to separate signals of different artefacts from brain signals given 8 sensor outputs(S) of a patient using ICA. However, the ICA algorithm also returns 8 source signals (S') which look the same as the input and hence I can't differentiate between signal and noise. How can I solve for this? Is there a preprocessing step to be done before feeding raw data to ICA?
I'm using Python and its packages for the same.
B. Of the 8 sensor outputs(S) I suspect some of them may be dependent causing the problem in (A). I suspect even in EEG recordings, a similar issue arises that electrode signals may be dependent. How can I solve for these dependent sources?
Relevant answer
Answer
What is your non-gaussianity method and which pre-whitening method do you use?
And more important question, can you assume that these sources were statistically independent?
  • asked a question related to ICA
Question
4 answers
Hello,
We are currently working on an ischemic stroke model on mice using intraluminal monofilament occlusion (using Koizumi's method). We use laser doppler to confirm the decrease in cerebral blood flow during MCA occlusion, and then the increase in cerebral blood flow during reperfusion. My concern is this: how can we determine if a resulting infarct as seen by TTC/cresyl violet staining is due to transient occlusion of the MCA (typically for 1 hour), or if it is due to permanent occlusion that may have developed post reperfusion (development of a thrombus due to arterial wall damage during surgery). We became aware of this issue after performing MRA on one of our mice post occlusion and then 24 hours post reperfusion; the MCA and part of the ICA had minimal collateral blood flow in the first image, then absolutely no blood flow in the 24 hour image. We could use laser doppler again to visualize cerebral blood flow just before animal sacrifice at 24 hours post reperfusion, however, it would be almost impossible to place the doppler probe in the exact same position as the initial recording. Any suggestions? Thank you very much.
Relevant answer
Answer
Hi Kristina
you can use PET scan to confirm your data.
The sensitivity and specifity of PET scan is high for this project.
I’m working on similar projects in neuroscience. I hope you follow me from this moment to be familiar with my articles and projects.
good luck
  • asked a question related to ICA
Question
8 answers
Hey,
No error message, no warning. The desktop is simply restarting itself... the script is working fine on my laptop. Windows' log says nothing about the crash event.
I tried several combinations of matlab (2014a, 2017b) and eeglab (12, 14...). This is a win10 x64 operation system (similar to my laptop).
"EEG = pop_runica(EEG, 'extended',1,'interupt','on','icatype','runica'); "
Relevant answer
Answer
Hi Kornél,
Sounds like like it could be a memory-related issue, but it's odd to see a whole system crash. Guessing you've already tried giving Matlab more heap space? Is it specific to the runica routine? Do the CPU temps spike while runica runs (bad fan perhaps?).
Just for test purposes (Artoni, Delorme, Makeig, 2018), what happens if you run runica with an initial pca decomposition (i.e. "EEG = pop_runica(EEG, 'extended',1,'interupt','on','icatype','runica','pca', [half/quarter number of actual electrodes here]); ". If everything still crashes, like Stephen said, the EEGlab mailing list is probably the best go-to place. Best of luck!
  • asked a question related to ICA
Question
7 answers
Since Biological Parametric Mapping (BPM) is no more supported, I would like to find another tool to combine several brain images acquired for each subjects or different groups.
I saw fusion ICA but it is not convenient for my data....
any other idea ?
Thank you
Isabelle
Relevant answer
Answer
If you want more flexibility in your ability to combine modalities, I would suggest extracting your values from the voxels of interest and treating them like any other variable in traditional statistical packages.
If you provide more information about exactly what sort of combined data analysis you're trying to do, we can answer more specifically.
  • asked a question related to ICA
Question
3 answers
I've got two raw data sets that has been gotten using OpenBCI.
I'm trying to understand the data from the channel spectra and maps(spectopo).
And, what does running ICA do? The components and the graphs, what do they mean?
Relevant answer
Answer
The files contain raw EEG data recorded with 250Hz sampling rate. So the ICA will process the data to eliminate any noise. Depending on your analysis and the study, the graphs will show the channel locations relative to the area of interest. Hope this helps.
  • asked a question related to ICA
Question
6 answers
I want to epoch the continuous EEG data of 8 minutes long into non overlapping epoch window in EEGLAB without any event information .When I select the 'Extracting Epoch " the GUI will pop up to set the Event information first .After epoch of the data I want to run ICA on the data to get the clean data .Kindly help me in this regard
Relevant answer
Answer
Hello,
I think you just need the eeg_regepochs() function to split the data into epochs of a specified lenght. see details here: https://github.com/jukka/eeglab/blob/master/functions/miscfunc/eeg_regepochs.m
  • asked a question related to ICA
Question
2 answers
I have obtaind 9 group ICs using GIFT. Now I want to extract each ICs of each subject using group ICs.
Subject matrx(contains voxels of each subject)=Mixing matrix(contains weight of each subject's ICA for each group component)*source matrix (voxels of each component)
  1. Can I multiply above weights of each subjects by each group components to obtain each subject's independent component?
  2. Or can anyone provide a script for this?
  3. Which method is more common?
Relevant answer
Answer
Dear Ediri,
every Individual ICA is calculated before group ICA, Gift sortware is a good tool for calculate and view components, but I have not experiece with that, There are good tutorials wich explain to use and set automatic scripts analisys. https://m.youtube.com/watch?v=q86nh8qZ228
The most common tools for ICA are fsl
And
CONN.
king regards
  • asked a question related to ICA
Question
2 answers
I did group ICA of 18 Subjects resting state fmri data now i want to regress the ICA output with smith 10 networks, so how do i design glm for this.
Relevant answer
Answer
It depends on what you are looking for. You can do the standard spatial ICA - fastest way will be to check correlation coefficient with known networks templates, you can also perform the temporal ICA and use Functional Network Connectivity Toobox to find temporal relatioships between components.
Good luck!
  • asked a question related to ICA
Question
4 answers
Currently ICA is applied for functional data. But...I feel this could be little debating.
Relevant answer
Answer
Hi Wasana,
That is an interesting idea. I can't see a major issue with it, and it has been used with diffusion tensor imaging.
I would suggest contacting Marcelo Febo at University of Florida for the opinion of an expert in a variety of brain imaging techniques.
  • asked a question related to ICA
Question
6 answers
This an example paper
Relevant answer
Answer
Hi Wellington Pinheiro Dos Santos . Thanks for your reply. I read in a book about PCA. The book is 'Statistical Analysis of Noise in MRI Modeling, Filtering and Estimation by
Santiago Aja-Fernández Gonzalo Vegas-Sánchez-Ferrero'
I want to write a review paper just about MRI denoising methods in transform domain.I have searched a lot but did not get a fixed opinion about methods in transform domain.
  • asked a question related to ICA
Question
11 answers
Hi all,
I would like to correlate several resting-state fMRI components extracted by ICA (using GIFT) with different behavioral measures related to my participants (as tests or questionnaires). Does anyone know the correct procedure?
Thank you in advance,
Edoardo
Relevant answer
Hi,
ICA is a nice method because the data analyst does not have to interfere with the analysis process, e.g. previously assuming a model (which could introduce severe bias, since the beginning). Therefore, it seems a bit contradictory to use ICA and then trying to correlate with whatever it is.
What you may do, in order to preserve the model-free strategy, is to compare two ICA sets acquired in two different conditions (e.g. with therapeutic vs. placebo, or male vs. female, or old vs. young). As each IC has its own timecourse, you just have to compare it with all the ICs’ timecourse of the other set, using any suitable statistic. Of course, if you start to have more conditions, the problem becomes untreatable very easily.
Another method is to use machine learning (e.g. artificial neural networks) for classification purposes. You just have to split your targets into categories. But in this case, for the training stage, you must know which ICs pattern correspond to each category. After building the model, you may start to make predictions (guessing targets) in order to assess if the model is acceptable (or not). You may find here some work that I did with this strategy:
I also used ICA with GLM here
You use the ICs’ timecourses as independent variables, but you should model the dependent variable with some sort of stimulation. I’m afraid that resting state isn’t a suitable paradigm for this.
Hope this helps. Regards.
  • asked a question related to ICA
Question
2 answers
Let's say I have 3 signals of specific brain areas that are linearly mixed, which I recorded with 3 electrodes plus a reference ground. I used FASTICA to separate the signals and I get 3 new separated signals. However I don't know which signal corresponds to each brain area, since the waveform resemblance with the original measurements is not obvious. I tried calculating the absolute value of the correlation between separated and measured signals to find which combination had a maximum correlation and this way classify the separated signals into their correct measured signal, however this does not always work. Does anyone have a suggestion or know another approach to solving this?
Relevant answer
Answer
ICA separates source signals by maximizing statistical independence using higher orders of the mean and some cost function. You can sort your ICA components by kurtosis, skewness to match them up. The munkres algorithm is also used to match components.. see:
there is also an implementation on nitrc for masked ica with BOLD data.
  • asked a question related to ICA
Question
11 answers
I doing signal processing and classification. May i know if i have 4 input signal, if there i combine 4 signal in one? or i need used something like ICA? or i add all 4 signal together before proceed with cnn? Or is that any other method i can do for multiple input of CNN.
Relevant answer
Answer
You can join all 4 to one signal or perform ICA. You can perform all kinds of transformations on the signal , depends on the signals type (is it audio, images electrical readings etc.) you may get better or worse results.
  • asked a question related to ICA
Question
6 answers
I did separate ICA analysis on 2 gropus (gift ICA), found components having highest correlations with DMN.
Should I next select the components and do one sample t-test in SPM on both groups separately (for thresholding) and then compare the results with two sample t-test?
I will be very thankful if anyone can help me.
Relevant answer
Answer
ICA exhibits several properties that make it challenging to perform a direct group comparison:
1) recomputing ICA on the same data does not necessarily yield identical results
2) ICA is an unsupervised method that does itself not perform a form of statistical inference - incorporating another criterion of optimality is typically helpful
3) there is hence no single best way to obtain a formal notion of group differences in ICA, which can be addressed by a number of ad-hoc procedure
-> a practical avenue can be to use the ICA projections for out-of-sample predictions in independent data
  • asked a question related to ICA
Question
3 answers
I am using FastICA to unmix some measurements. Although my initial measurements are positive, I obtain independent components that has positive and negative signals.
When I made a search about it I only found a forum entry:
"...The ICA components are estimated "up" to a permutation, a scale and a sign factor. That means the ICA components are NOT intrisically ordered, are "convetionally" scaled to spatial z-scores and the sign does not carry any "absolute" information"
Can I do some operations on the results like taking the absolute value or shifting everything up by subtracting the minimum? In the case of shifting, the whole background (which supposed to be zero) takes a positive value.
Is there a way to force output to be positive values only? Or why those polarity occurs. I would be glad if someone could help me to interpret the results.
Relevant answer
Answer
HI Mustafa,
The negative values of the ICs happen because ICA weight matrix has undetermined polarity (sign). If your problem physically ensures only positive values, taking the absolute values of the ICs will be enough
Hope it helps
Regards
  • asked a question related to ICA
Question
4 answers
in some parts of applying ICA to an EEG data we have blind source seperation problem. how can we solve this problem?
Relevant answer
Answer
First of all you must know the mixing mechanism of the signals. Then apply algorithm related to that mixing model.
  • asked a question related to ICA
Question
1 answer
Hello 
I have been trying MELODIC group analysis and get error after completing and didn't get final group ICA:
child process exited abnormally
> while executing
> "fsl:exec "${FSLDIR}/bin/feat ${fsfroot}.fsf -D $FD -I $session -prestats" -b $howlong -h $ID -N feat2_pre -l logs "
> invoked from within
> "if { $done_something == 0 } {
>
> if { ! $fmri(inmelodic) } {
> if { $fmri(level) == 1 } {
> for { set session 1 } { $session <= $fmri(mult..."
> (file "/usr/local/fsl/bin/feat" line 387)
 
 
When the MELODIC start, then I get this message
Feat main script
/bin/cp /tmp/feat_0y46We.fsf design.fsf
mkdir .files;cp /usr/share/fsl/5.0/doc/fsl.css .files;cp -r
/usr/share/fsl/5.0/doc/images .files/images
Please let me know how to reslove this problem.
Thanks and regards
Sadhana
Relevant answer
Answer
You can try to reinstall the software and redefine working path in terminal.
If the program works well, there are online tutorials very helpful, like this:
Best regards
Riccardo
  • asked a question related to ICA
Question
8 answers
I do the data analysis in EEGLAB and want to do ICA to remove the artifact. As the EEG data is recorded in .bdf, so I have two questions:
1. There are no reference in the recorded data and when I import EEG data into EEGLAB, I had to chose one channel as the ref, this may cause the rank-deficient;
2. I had to deal with the bad channels before I do the ICA, but I find if I do the interpolation before ICA, the rank of the data also reduced (interpolate one channel make the rank-1).
As rank deficient will make mistake for ICA result, so does someone konw how to deal with these two problems? Thanks.
Relevant answer
Answer
Hi wenwen,
1, I think you can use the average reference after loading the data into EEGLAB.
2. After you remove the bad channels, you can only decompose your data into components as many as good channels. After that, you can bring your bad channels back by interpolation.
Ping
  • asked a question related to ICA
Question
6 answers
Once a professor mentioned us that certain algorithms for dimensionality reduction and visualization work for certain types of data. Many of us know that, for example, PCA, ICA, t-SNE, autoencoders etc work well in continuous data type. But what if you have a combination of categorical, categorical-ordinal data, binary and continuos data... what are the correct steps and algorithm to manage such combination of information or by separate (categorical data, binary, categorical-ordinal data)?
Do the previous algorithms work on well on these type of mix data types? The one-hot encoding is enough to threat the categorical data and then use the respective transformations? How we treat categorical-ordinal data... do we apply one-hot encoding and then transformation techniques or make the assumption as it is a continuous data?
Relevant answer
Answer
Dear Manuel Martinez, from my perspective, most useful methods such as PCA, ICA, autoencoders etc relies upon a fundamental notion of distance between patterns. Note that these methods may even yield discrepant results depending on the chosen distance (Euclidean, Mahalanobis, Manhattan, etc.), which corroborates the idea that distance is a more fundamental concept than the pattern representation itself.
 If you agree with this point, then you will also see that the most important aspect of your question would not be if patterns are represented as numerical or categorical, but if you can define meaningful distances between them (not necessarily the distances imposed by a given numerical representation, for it can be meaningless, sometimes).
Accordingly, what I suggest is two-fold:
a) Define a clear and meaningful measures of distance between patterns, no matter if it is categorical, numerical or even heterogeneous. Clearly, the better your distance measure, the more it should reflect the phenomenon behind collected data.
b) Use Multidimensional Scaling to find a suitable metric space for all your data (i.e. a single and meaningful metric space were all  kinds of patterns can be translated into a real-valued vector).
After these 'pattern translation' steps, all of-the-shelf methods can be easily used.
Clearly, step (a) is the most important and difficult one. But it may be surprisingly simple! For a practical and successful example of step (a), please have a look on the word2vec strategy, where the concept of 'context' is the key to the conception of a kind of meaningful distance between words (instead of comparing words themselves).
  • asked a question related to ICA
Question
3 answers
I am looking for a library to be integrated in LabWINDOWS CVI.
Relevant answer
  • asked a question related to ICA
Question
3 answers
Hi experts
I want to compare ICA methods using ICASSO in Gift toolbox for rsfMRI. How can I find out what parameter settings are appropriate for each method?
I run 6 methods with same parameter setting and as the result, some of them were good some were bad and some of them were irrational. I attached a picture that shows some of my results. Please help me find out.
regards
Relevant answer
Answer
Basically what ICASSO's  output parameters tell you is how good is the clustering of extracted independent components (ICs) after running ICA repeatedly. Good clustering indicates stable decomposition. For example, you can see in your figures that InfoMax ICA clusters are tight and well separated, which means extracted ICs were quite similar across different ICA runs => stable decomposition (check other output parameters as well for reliable conclusions).
In other words, ICASSO won't directly tell you what ICA parameters are appropriate or even which algorithm is appropriate. But it will tell how ICA convergences and how stable decomposition is, so you can indirectly use it for tweaking parameters, for example by running ICA with different parameters and by examining ICASSO output.
  • asked a question related to ICA
Question
3 answers
dual_regression is a powerful script for extracting time series data for each ICA maps in RSN analysis. Can we use dual_regression for task-based fMRI studies? Another thing can we use a special cortical mask instead of ICA maps files. Any ideas
Relevant answer
Answer
Dear Carlos,
One of my problems when using ICA is how to deal with maps? You know the ICA maps may distributed on all brain spheres, but my goal from using ICA is identifying small specific ROIs. they are some thing between voxels and anatomical regions. Is that feasible by ICA?
  • asked a question related to ICA
Question
16 answers
I have an EEG data set which is about 5 minutes long for each subject. I want to detect and correct existing artifacts using ICA approach. I can apply this method on the whole data of each person or first epoch this data and then apply ICA on each epoch separately. Which one is more accurate? 
Relevant answer
Answer
ICA algorithms typically take sample*channel matrices. So epoching itself will not have any effect on the results of the algorithm because the data is re-concatenated in most packages/programs (e.g. EEGLAB)  before submission to the algorithm. 
Removing data samples in the form of entire epochs will have an effect on the results of ICA algorithms. Especially if those epochs contain large, unique movement artifacts. ICA algorithms will typically place unique movement artifacts in single components.
It would probably be best to submit all epochs to the algorithm unless there are some egregious movement artifacts in the recording. No EEG will be able to be recovered during periods of large movement artifact.
Was this helpful? I've attached a book chapter for reference. Section 6 contains the relevant discussion. 
  • asked a question related to ICA
Question
3 answers
Hi all,
I used intependent component analysis to sepereate speech from noise.i used 2 microphones and two sources (speech and noise) . I got improvement in SNR when I used ICA but when I used ICA in speaker verification . I got bad performance compared with noisy test signal. Does any one used ICA to remove noise and used the enhanced speech after ICA in speaker verification?
Relevant answer
Answer
I did some experiments on noisy speaker recognition using speech enhancement with non-negative matrix factorization (check FASST toolkit)
The approach works well, but only if you apply it for both noisy train and test data. To my knowledge, any speech separation algorithm leaves some residual noise, and consequently, there is still a mismatch between clean speech and the enhanced one. If your training data are clen, I'd suggest to apply artificial noising techniques.
It might also depend on your classifier. For me the best was obtained with i-vector extractor and plda classifier trained on noisy data enhanced with NMF approach
  • asked a question related to ICA
Question
4 answers
 Can rythm isolation ( mu and beta ) be done on