[Question] Can Unfold.jl be used to seperate evoked vs induced responses

[duodenum96]

This is a question out of curiosity: In my own research I encountered a problem on seperating evoked vs induced activity. The conventional way to perform this is subtract evoked ERP (average over trials) from individual trials and call it induced since ERP will reflect phase-locked responses to event and induced will consist of the remainder. But if there is jitter between evoked responses to stimulus (in terms of latency for example) then this operation fails (this fieldtrip tutorial gives a good explanation: https://www.fieldtriptoolbox.org/faq/spectral/evoked_vs_induced/). I was curious whether the mixed-effect approach of Unfold.jl can address this problem. This can be a seperate research project in and by itself.

[behinger]

Thanks for this suggestion!!

I think there are two aspects. 1) Can Unfold be used to model jittered / morphed stimuli. The answer is yes, but no ;). In principle, if you know the ERP shape, you could generate jittered/squeezed versions of that shape, run a PCA on them, and include them as predictors, very similar to a canonical BOLD + derivatives in fMRI research - except that we dnt know the BOLD, ergo we have to estimate it first. This is the approach implemented by https://pubmed.ncbi.nlm.nih.gov/24110660/ et al. I recently had a student implement something like this, but we noticed that the approach is a bit limited given that you have to define yourself, when a component of an ERP is "over", in order to stretch it properly. Most likely there are other ways to find that canonical kernel, and thenyou could add it to a unfold-type of analysis. This would also allow you to combine it with overlap with other events, potentially in a RIDE-style iterative approach; but o.c. this is two papers ahead of where we are right now

2. can you maybe model the induced & evoked part directly? On something related @geigermartin is working on in a sideproject. He tries to model a complex EEG signal, e.g. from a hilbert transform, to investigate overlapping potentials in TimeFrequency analyses. The problem there is, that the induced part will have a per-trial differing phase; thus how to get unfold to fit the phase but not the amplitude; tricky!

Is this what you had in mind?

Best, Benedikt

[duodenum96]

Ah, I see. What I had in mind was, very handwavily:

evoked_{trial} = f(amplitude, latency)

Then there might be one fixed effect for average amplitude and latency and random effects for each trial. Induced power would be ERP_{trial} - evoked_{trial}. But given your point 1, it doesn't seem very realistic AFAICT. Thanks for the answer!