[Eeglablist] Practically reproducible results with AMICA

Wed Oct 7 12:41:13 PDT 2026

Hi,

I don't want to add more compute time, but I'd be curious to see how these results are affected by filtering (we all know highpass is crucial and gives very different results if you remove low-freq drift or not, that's in Winkler et al. 2015, but I'd be curious about lowpass filtering between 30 hz vs lowpass 45 hz vs just a notch filter at power line freq), and different sample rates (e.g. 250 vs 500 hz to keep it simple). I would suspect the reproducibility would increase with lower filter cutoff (e.g. 30 hz since otherwise you get more line noise related components) and higher sample rates. 

Good question about data cleaning, that likely also affects the convergence of those types of results. I'd recommend ensuring always highpass filter at 1 Hz and a conservative ASR cleaning (e.g. at a threshold around 100 (depending on the data quality, the most important is that it doesnt remove good brain activity like alpha or ocular artifacts so they are separated properly). Of course before ASR and ICA, make sure to remove bad channels (e.g., with clean_rawdata plugin or clean_channels function directly), and make sure to input the effective data rank when running ICA (see Kim et al. 2023 if you haven't already), I can't remember if that is done automatically under the hood or not, but I do remember the input option is available for sure. Jason can confirm. 

I would also like to see comparison of results with different re-referencing schemes (CAR, REST, Surface Laplacian) but with 26 channels, you're pretty limited in options here. I'd recommend doing this project on at least 64 electrodes dataset (and probably compare also resting state with other task-based datasets, where reproducibility will probably vary differently). I would suspect CAR and REST may converge similarly (although I'd love to see the results) and Surface Laplacian to be less reproducible (you could drop that one to save compute time since most people don't do that much anymore). 

Good luck!

Cedric

Sent with Proton Mail secure email.

On Wednesday, October 7th, 2026 at 12:18 AM, Grace Harvie via eeglablist <eeglablist at sccn.ucsd.edu> wrote:

> Hi Jason,
> 
> Thanks for the helpful input!
> 
> I was using the default iterations of 2000, but experimented with iterations of up to 15000, as I noticed that the algorithm would hit the maximum number of iterations. When set to 15000 iterations, the first run performed approximately 12000 iterations before exiting because of the likelihood increase (<1e-9), while the second run performed 7407 iterations before exiting for the same reason.
> 
> For that last run of 7407 iterations, LL = -3.1371323664, however I did notice on both of the 15000 iteration runs, the nd didn't seem to decrease, hovering between 2.8718e-5 and 8.0337e-6 for the last 2000 iterations of that last run, like you describe.
> 
> I should probably describing the data I'm working with - 2 minute resting state recording from 26 cephalic channels sampled at 500Hz, however the participants were paediatric/adolescent ADHD patients, so there are sometimes quite significant movement/muscle artifacts present in the recordings. I've tried to strike a balance between adequate cleaning (highpass filtering, cleanline, ASR) prior to allow high quality ICA decomposition and not over preprocessing the data (As per Arno's 2023 article, "EEG is better left alone", https://urldefense.com/v3/__https://doi.org/10.1038/s41598-023-27528-0__;!!Mih3wA!HRj9wPvczBfozBYg895WdVhu1e5D2_L1MLwYWRWA1aaku0I9KQ-KDM_aJHGiVucuVMBVj8HnJcdsbmZh3dNvG-UK1PIeeZCQ$ ).
> 
> It sounds like the quality of the data might be prohibiting a high quality decomposition, which concerns me, because the EEG recording I've been testing this on is one of the cleaner in the dataset.
> 
> There are definitely many common component maps across the multiple runs, so I might play around with matcorr.
> 
> Many thanks,
> Grace Harvie
> 
> PhD Candidate, Faculty of Medicine and Health, University of Sydney
> 
> Research Assistant, Brain Dynamics Centre, Westmead Institute for Medical Research
> 
> 176 Hawkesbury Road | PO Box 412 | Westmead NSW 2145 Australia
> ghar8070 at uni.sydney.edu.au
> westmeadinstitute.org.au
> 
> 
> This email is private and confidential to the intended recipient. If you are not the intended recipient, please do not copy it, circulate it or take any action in reliance on it. Kindly notify me that it has been misdirected and then delete it. Thank you.
> 
> ________________________________
> From: Jason Palmer <japalmer29 at gmail.com>
> Sent: 07 October 2026 07:38
> To: Grace Harvie <ghar8070 at uni.sydney.edu.au>; eeglablist at sccn.ucsd.edu <eeglablist at sccn.ucsd.edu>
> Subject: RE: [Eeglablist] Practically reproducible results with AMICA
> 
> Also the "nd" in the output is the average weight change and should be less
> than 1e-5 or so. The issue is whether the algorithm has fully converged.
> Ideally the LL should be increasing by less than some small number like
> 1e-9, and the weight change should go to zero, or nd < 1e-7 e.g. Sometimes
> the nd may actually only get to 1e-5 or so and "hover" if there is noise
> causing an unstable optimum.
> 
> The components are arranged in variance order, where v = sum(EEG.icawinv.^2)
> .* var(EEG.icaact(:,:)'). If the IC variances are close the order may change
> while the component (map) doesn't change. The matcorr function can match the
> columns of the icawinv map matrices.
> 
> Best,
> Jason
> 
> -----Original Message-----
> From: Jason Palmer <japalmer29 at gmail.com>
> Sent: Tuesday, October 6, 2026 1:33 PM
> To: 'Grace Harvie' <ghar8070 at uni.sydney.edu.au>; eeglablist at sccn.ucsd.edu
> Subject: RE: [Eeglablist] Practically reproducible results with AMICA
> 
> Hi Grace,
> 
> The number of iterations should be the main factor. Are you using the
> default 2000? I usually do 2500, but if you are testing reproducibility,
> then I would recommend 5000 max iterations. (It's a good idea to set
> writestep or Write Out iterations to 500 for speed purposes, and set
> 'do_opt_block' or optimize block size to checked. And the maxthreads to
> whatever your system can handle, like 4 if you have 8 cores e.g. This is all
> just for speed.) Also, you can check the final value value of LL, which is
> what AMICA is maximizing.
> 
> Best,
> Jason
> 
> -----Original Message-----
> From: eeglablist <eeglablist-bounces at sccn.ucsd.edu> On Behalf Of Grace
> Harvie via eeglablist
> Sent: Monday, October 5, 2026 7:46 PM
> To: eeglablist at sccn.ucsd.edu
> Subject: [Eeglablist] Practically reproducible results with AMICA
> 
> Hi All,
> 
> I'm back with another AMICA question - this time regarding obtaining
> practically reproducible decompositions with AMICA.
> 
> I was able to obtain consistent and reproducible results with extended
> INFOMAX by utilising the parameters described by Makoto's here:
> https://urldefense.com/v3/__https://url.au.m.mimecastprotect.com/s/NXddCBNqjlCAJWVG7fzfyjs25FIy?domain=eeglab.ucsd.edu__;!!Mih3wA!HRj9wPvczBfozBYg895WdVhu1e5D2_L1MLwYWRWA1aaku0I9KQ-KDM_aJHGiVucuVMBVj8HnJcdsbmZh3dNvG-UK1Mj2Oevp$
> ctically_reproducible_ICA_results_(09/26/2022_added,_05/08/2024_updated)
> 
> While Makoto comments that the default parameters for AMICA are much more
> refined and seem to produce similarly stable, replicable results, I have
> observed variable results over multiple runs with AMICA - please see the top
> 7 ICs for ten AMICA runs on the same dataset here:
> https://urldefense.com/v3/__https://url.au.m.mimecastprotect.com/s/6Z8yCD1vlpTMrYB95t5hG6sjClvb?domain=urldefense.com__;!!Mih3wA!HRj9wPvczBfozBYg895WdVhu1e5D2_L1MLwYWRWA1aaku0I9KQ-KDM_aJHGiVucuVMBVj8HnJcdsbmZh3dNvG-UK1D1WXMYn$
> GData/blob/main/AMICADecompositionReproducibility.png__;!!Mih3wA!ECb3eiORNCe
> yUP1u_NEtlhsE66TmRqljEcyRtGZyp6WqAlxLU16gwSe69jtFfeZAUuOzjacBKM-x3Q40quQJDJc
> 52OiFBKxr$
> 
> Additionally, on one run (Run 8), I received a warning about the ICA
> activities and weight matrix not matching as soon as AMICA finished. I'm not
> sure why I got that warning, especially since EEG.icaact and EEG.icaweights
> data structures appear to each have the correct number of values. I'm also
> not sure if it's relevant to the issue at hand, but I thought it was
> probably a good idea to mention it anyway.
> 
> Since Makoto showed that INFOMAX reproducibility could be improved by
> increasing the number of iterations and lowering the learning rate, I tried
> to similarly tweak these parameters in AMICA, however it didn't seem to
> help.
> 
> Does anyone have any recommendations to help improve reproducibility of
> AMICA results?
> 
> Many thanks,
> Grace Harvie
> 
> PhD Candidate, Faculty of Medicine and Health, University of Sydney
> 
> Research Assistant, Brain Dynamics Centre, Westmead Institute for Medical
> Research
> 
> 176 Hawkesbury Road | PO Box 412 | Westmead NSW 2145 Australia
> ghar8070 at uni.sydney.edu.au westmeadinstitute.org.au
> 
> 
> This email is private and confidential to the intended recipient. If you are
> not the intended recipient, please do not copy it, circulate it or take any
> action in reliance on it. Kindly notify me that it has been misdirected and
> then delete it. Thank you.
> 
> _______________________________________________
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> or visit https://urldefense.com/v3/__https://url.au.m.mimecastprotect.com/s/3Tu4CE8wmrtno4W73upiG9s7SMf6?domain=sccn.ucsd.edu__;!!Mih3wA!HRj9wPvczBfozBYg895WdVhu1e5D2_L1MLwYWRWA1aaku0I9KQ-KDM_aJHGiVucuVMBVj8HnJcdsbmZh3dNvG-UK1NB_BbtX$  .
> 
> 
> _______________________________________________
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu or visit https://sccn.ucsd.edu/mailman/listinfo/eeglablist .
>