Validity of Sliding-Window FCD as a Fitting Target

[Szymon Tyras]

Hello everyone,

I am using the E-I RWW model with a fitting target of the FCD (BOLD) matrix, calculated using the sliding-window method. I am correlating some behavioral features with simulation-derived data and am getting interesting results.

However, I recently read this paper and started wondering if it can still be considered a valid fitting target: https://www.biorxiv.org/content/10.1101/2023.10.06.561221v2.full.

Specifically, in my case, on one hand, it is still a very popular method (also in the whole-brain modeling field). My results are reproducible when I use different parcellations, different window sizes, and different filtering methods. Additionally, when I add framewise displacement as a covariate, my results become even clearer. On the other hand, when I tested it with different fitting metrics like static FC or phaseFCD, I did not see the same results (sometimes even showing opposing trends but less significant). While it is possible that other metrics simply do not relate or relate differently to my behavioral features, it could also indicate some errors introduced by the sliding-window method.

So, I am generally interested in the community's view on using sliding-window FCD as a fitting metric and whether you have observed different results when fitting to, say, phaseFCD.

Thank you so much for your insights.

Best regards, 

Szymon

[Randy McIntosh]

Hi Szymon,

 

Glad you’re getting interesting results.  I don’t know if anyone has compared model parameters derived from fitting window-FCD vs phase-FCD.   We’re working to optimize models based on more than one data feature to get ensure we’re not biased to residual noise in the data (which is what I think Martin’s paper suggests).  This includes FCD variance, FCD distance between empirical and sim, and staticFC.   What feature of FCD do you fit?  Also, you can do the checks that Martin suggests in his paper to ensure your FCD measures are not contaminated.

--
You received this message because you are subscribed to the Google Groups "TVB Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tvb-users+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/76bdf07e-571f-441b-ac34-879de7d6ce13n%40googlegroups.com.

[Szymon Tyras]

Hello Randy,  

Thank you so much for your response. I was fitting to FCD distance using K-S statistics. After checking, I found that the results were entirely driven by the global signal. Therefore, I believe that this, combined with the absence of those results in different fittings, contradicts my initial observations.

Best regards,

Szymon

[Daniele Marinazzo]

The global signal is the coarsest region of interest, it is very correlated with systemic fluctuations as well as "noise", but it contains the whole brain after all, even if some (even prominent) colleagues sell with hype and (I hope fake) surprise the fact that "there's signal in the noise", "global signal is correlated with behavior" and things like that.

So you could be fitting neural activity after all, albeit in a non region-specific way. You could try global signal regression after correcting for blood arrival time (https://www.frontiersin.org/articles/10.3389/fnhum.2016.00311/full) and see what happens.

To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/fbcaed72-4adc-4a1e-9763-ee813a065af6n%40googlegroups.com.

[Szymon Tyras]

Dear Daniele,  

My sincere apologies for the delayed response—I must have somehow overlooked your message. Thank you for your suggestion; Indeed, I may give it a shot by exploring those parts of the GBS results.