Impulse response function and deconvolution for gPPI analysis with fast TR

Hello AFNI team,

I'm completing gPPI analyses on several experiments with different acquisition parameters and am using the steps in this guide: Context-Dependent Correlation Analysis or Generalized PPI | afni.nimh.nih.gov

Should there be specific considerations when creating the impulse response function or completing the deconvolution on the seed time series if the TR is quite fast? The original dataset I analyzed had a TR of 2s although I'm trying to replicate the analysis on data with a TR of 0.8s and want to ensure I've created the IRF and interaction regressors correctly.

I created the impulse response function & applied it using:
waver -dt 0.8 -GAM -inline 1@1 > GammaHR.1D
3dTfitter -RHS Seed _ts.1D -FALTUNG GammaHR.1D Seed_Neur 012 0

This is the waver command when creating the interaction regressor
waver -GAM -peak 1 -TR 0.8 -input Inter_stim.1D -numout #TRs > Inter_A.1D

Hello AFNI team,

I've been a bit stuck on the gPPI analysis for the study with the TR=0.8s. The studies are the same experimental design but with different stimuli and the difference in TR (0.8s versus 2s). I've not found any significant findings when trying to replicate the original study. Would there be any special considerations for conducting the gPPI analysis with a faster TR?

Thank you,
Catherine

Catherine,

The disparities observed between your two datasets could shed light on the degree of robustness of the PPI approach. It's established a priori that the shape of the hemodynamic response varies substantially across the brain. Consequently, expecting a canonical impulse response curve to reasonably capture nuanced fluctuations across trials might be unrealistic.

One viable alternative is to estimate the hemodynamic response without presuming its shape in advance and, concurrently, derive inferences about cross-region relationships. Here is an approach called CPCA that is based on principle component analysis.

Gang

Hi Catherine and Gang,

Thanks Gang for mentioning fMRI-CPCA. CPCA is a more general statistical method. fMRI-CPCA uses a FIR model. A FIR model does not really change with shorter TR, you just need to estimate about 20-30 seconds of task-induced BOLD changes after stimulus presentation, averaged over trials. So with a TR of 2 that is 10-15 FIR time bins. With a TR of .8, that would be 25-38 FIR time bins.

Any questions? toddswoodward@gmail.com

Hello,

I plan on exploring FMRI-CPCA using FIR model more in the future (thank you for the offline help Todd!)

In the meantime I was thinking more about how to conceptually interpret the connectivity regressors in relation to the task activity regressors. My understanding is that gPPI effects are independent of the task-related activation only effects. I'm specifically interested in 1) activity changes in an ROI across 4 task conditions and 2) functional connectivity changes using the same ROI across the same 4 task conditions. The same 4 conditions and ROI are used in both experiments. I would predict a similar pattern of changes with the ROI for both the task-related activation and functional connectivity (and for both experiments).

For experiment 1, I did not see any significant changes in the brain activity across the 4 task conditions in the ROI. Yet I did see significant changes in the functional connectivity changes between the ROI & other regions across those conditions.

For experiment 2, I did not see significant changes in the functional connectivity across the task conditions but did see significant changes in the brain activity.

Since I'm expecting the task-related activity and functional connectivity to both show the same pattern of changes across the task conditions, I was considering that significant changes in brain activity would reduce the likelihood of detecting changes in functional connectivity (or vice versa).

Both regressor types use the same brain signal yet with differing de/convolutions and also are included as separate regressors in the GLM. Are there any general thoughts on the shared variance between the task and gPPI regressors?

It might be more accurate to state that PPI is designed to capture the interaction between trial-level effects (relative to the condition-level average effect) and the seed region.

For experiment 1, I did not see any significant changes in the brain activity across the 4 task conditions in the ROI. Yet I did see significant changes in the functional connectivity changes between the ROI & other regions across those conditions.

For experiment 2, I did not see significant changes in the functional connectivity across the task conditions but did see significant changes in the brain activity.

Do these observations still hold even when you substantially lower your significance stringency?

Gang

1 Like

Hello Gang,

Thank you for the clarification!

No, at a more liberal cluster threshold the expected pattern is there for the PPI effects in experiment 2. In experiment 1 I still only observe the significant PPI effects and not the condition average effects.

I still believe that the issue could be linked to the inflexibility of the assumed hemodynamic response function, but this is purely speculative. Of course, stochastic factors might also play a role.

Gang