glHMM application on multicenter data with two timepoints

[nooshinjava]

Hello, I am very interested in applying GLHMM to a large-scale longitudinal dataset including resting-state and task-based fMRI.
As I set up my pipeline, I would be grateful for your advice on several methodological and implementation aspects:
1- Modeling across conditions – Is it preferable to fit separate GLHMMs for each fMRI condition (rest and tasks), or to concatenate time series across conditions/sessions into one model per subject?
2- Longitudinal design – For baseline and follow-up data, would you recommend modeling sessions jointly (e.g., stacked time series), or fitting models separately and comparing results post-hoc?
3- Scanner and acquisition effects – The dataset spans multiple scanners. Should such effects be regressed out before GLHMM (e.g., standardized time series), or explicitly modeled in downstream analyses?
4- Number of states (K) – What strategy do you recommend for state selection? Should K be optimized separately per condition, or fixed across conditions for comparability?
5- Unequal scan lengths / missing sessions – How should these be handled to avoid bias in model estimation?
6- Task modeling – For task fMRI, would you suggest incorporating event labels per TR (e.g., condition-specific modulation) into the GLHMM, or is it preferable to model the dynamics without task regressors?
Thank you very much for your time and for developing this toolbox. Any guidance you can provide on best practices for applying GLHMM to complex longitudinal and multi-condition data would be greatly appreciated.
Bests, Nooshin

[LauraMasaracchia]

Hello, thank you for your interest in the toolbox. Below we try to answer to your questions concisely.
1-
That depends on your research question. If the question is to find differences between conditions, fitting it per condition is a more expressive model and therefore is more likely to capture differences. Fitting it for all conditions together makes the comparison easier to interpret (because state 1 for example is the same for all conditions).
2-
The same applies. We would maybe start by fitting them altogether and see if there’s any difference. In this case, differences can only exist in the temporal patterns of the state, but not on their spectral or spatial properties.
Alternatively, you can have a run with all sessions concatenated, and then use the resulting state time courses to initialize a different estimation per condition / time point.
3-
The more you can do to homogenize the data at the preprocessing stage, the better. But we would also include the scanner in downstream analyses (e.g. statistical testing) to be sure.
4-
An HMM with a higher number of states will probably fit better the data, but the estimate will be a bit less stable. For state selection you can for example use the free energy, and/or half-splits of the data and separate runs to check reproducibility. Also, we recommend running the HMM with different number of states and comparing the results.
5-
If you mean these differences are across conditions, one option is to match to the shortest
condition by shorten the data from the other conditions; and then do multiple
repetitions where each time you do the matching a bit differently.
6-
Both options are possible, and that depends on the research question.
Some examples that made use of the label are:
https://academic.oup.com/cercor/article/29/2/863/5232535
https://onlinelibrary.wiley.com/doi/pdf/10.1002/hbm.25835

We hope our responses helped you with your analyses. Feel free to reach out again, might you have more questions!