Fitting a model to data from the same subject on 2 distinct recording sessions

[NeuroLife]

Hi,

I was wondering if anyone had any insight about fitting resting state data from the same individual that were taken on different days (separated by weeks or months and taken in similar context).

Would there be any meaning to fitting them separately assuming that we use a generic structural connectome? My worry is that by doing so, we would not capture information about the individual and it would be equivalent to fitting 2 distinct people. So, I was thinking that maybe by fitting the model on the first day, then selecting specific parameters for which a variation over time makes sense (such as neural mass population coupling strength) and try to fit the data from the second day using those parameters by changing their value slightly from the values found for the first day we would reduce this loss of information (perhaps at the cost of goodness of fit).

If anyone has knowledge about this or could refer me to a source that may enlighten me on this subject it would be greatly appreciated.

Thank you for your time!

Dominic Boutet

[WOODMAN Michael]

hi,

These are good questions, for which there aren't clear answers.  The structure of the data suggests using a hierarchical Bayesian model, where you would pool parameters over different data sets, but also allow for per data set parameter values.  The pooled parameters can place priors on the per day parameters etc.  Usually there are several variations of the hierarchical structure and Bayesian model comparison can be used to choose the 'best' model. 

If that's of interest, please read more in the Stan User Guide to get ideas on structuring your overall model for your dataset.  TVB-style models can be implemented in Stan as part of your hierarchical model or in another probabilistic programming framework (e.g. TensorFlow probability).

cheers,

Marmaduke

---

From: tvb-...@googlegroups.com <tvb-...@googlegroups.com> on behalf of NeuroLife <domin...@hotmail.com>
Sent: Sunday, August 15, 2021 6:47:44 PM
To: TVB Users
Subject: [TVB] Fitting a model to data from the same subject on 2 distinct recording sessions

 

--
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/d20b53c4-cf6f-4ec6-9ad7-bbc1a7842d9bn%40googlegroups.com.

[NeuroLife]

Thank you again, this is very helpful!

Have a nice day!

Dominic Boutet

[Viktor Jirsa]

Hello,

The following may be also useful in this context: Melozzi et al PNAS 2019. Paper is attached.

Francesca (the first author) scanned 19 mice multiple times. Scan-rescan variability was about 65% when assessed with functional connectivity (FC). This metric has many issues but allows you to provide a first quantification of variability. To go beyond, she systematically modelled the dynamics with different connectomes (Allen, different surrogates, individual) and analysed the FC and other metrics, all compared against empirical functional data. You may find some of the approaches and results in there useful.

Best, Viktor

[NeuroLife]

Thank you, this is very interesting!

I may be a little out of my depth here, but would this imply that, when using the same SC to model both recordings and only fitting the local & global dynamics of the model to say spectral estimates (PSD), this variability may be involved in the relatively significant difference between the best fit for both recordings of the same individual? As if the difference in local dynamics would be a compensation (overfitting?). Also, to be more specific about the context, I am referring to MEG recordings.

Thank you for your time,

Have a nice day!

Dominic Boutet

[WOODMAN Michael]

hi

Yes best fits in two recordings reflect scan-rescan variability.  I don't think attributing it to overfitting per se is correct though, as it could be the estimator (PSD with short time window is a high variance estimator, better to use a windowed method) or actual variability in the brain.

cheers,

Marmaduke

---

From: tvb-...@googlegroups.com <tvb-...@googlegroups.com> on behalf of NeuroLife <domin...@hotmail.com>

Sent: Friday, August 20, 2021 3:30:48 PM
To: TVB Users
Subject: Re: [TVB] Fitting a model to data from the same subject on 2 distinct recording sessions

 

--
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/07821a0a-70eb-4e8e-a74b-3a67f41ae98fn%40googlegroups.com.

[Viktor Jirsa]

Hello again,

In this context, Saggio et al 2016 maybe helpful who used linear models to link second order statistics (FC, which can be linked to PSD via the Wiener–Khinchin theorem) to SC analytically. As this is analytical, you get a cheap and fast fitting and can estimate local vs global contributions efficiently. 

Best, Viktor

 

 

[NeuroLife]

Hi,

This has been very helpful!  Thank you both for taking the time to answer my questions.

Have a nice day!

Dominic Boutet