One of the things we've been busy with during the pandemic-induced data collection pause is reviewing the DMCC (Dual Mechanisms of Cognitive Control; more descriptions here and here) project GLMs: did we set them up and are we interpreting them optimally?
We decided that the models were not quite right, and are rerunning all of them. This post (and its successors) will describe how the original GLMs were specified, why we thought a change was needed, what the change is, and how the new results look. The posts are intended to summarize and clarify the logic and process for myself, but also for others, as I believe many people find modeling BOLD responses (here, with afni 3dDeconvolve TENTs) confusing.
I won't just start explaining the DMCC GLMs however, because they're complex: the DMCC has a mixed design (task trials embedded in long blocks), and we're estimating both event-related ("transient") and block-related ("sustained") responses. If you're not familiar with them, Petersen and Dubis (2012) review mixed block/event fMRI designs, and Figure 1 from Visscher, et. al (2003) (left) illustrates the idea. In addition to the event and block regressors we have "blockONandOFF" regressors in the DMCC, which are intended to capture the transition between the control and task periods (see Dosenbach et. al (2006)).
We fit many separate GLMs for the DMCC, some as controls and others to capture effects of interest. For these posts I'll describe one of the control analyses (the "ONs") to make it a bit more tractable. The ON GLMs include all three effect types (event (transient), block (sustained), blockONandOFF (block start and stop)), but do not distinguish between the trial types; all trials (events) are given the same label ("on"). The ON GLMs are a positive control; they should show areas with changes in activation between task and rest (30 seconds of fixation cross before and after task blocks). For example, there should be positive, HRF-type responses in visual areas, because all of our task trials include visual stimuli.
Below is an example of the estimated responses from our original ONs GLMs (generated by R knitr code similar to this). The GLMs were fit for every voxel individually, then averaged within the Schaefer, et al. (2018) parcels (400x7 version); these are results for four visual parcels and volumetric preprocessing. These are (robust) mean and SEM over 13 participants (DMCC13benchmark). The results have multiple columns both because there are four DMCC tasks (Axcpt (AX-CPT), Cuedts (Cued task-switching), Stern (Sternberg working memory), and Stroop (color-word variant)), and because of the mixed design, so we generate estimates for both event/transient (lines) and sustained (blue bars) effects (blockONandOFF are not included here).
The line graphs show the modeled response at each time point ("knot"). Note that the duration (and so number of estimated knots) for each task varies; e.g., Stroop has
the shortest tasks and Stern the longest. (We set the modeled response to trial length + 14 seconds.)
The double-peak response shape for Axcpt, Cuedts, and Stern is expected, as the trials have two stimuli slides separated by a 4 second ITI; the Stroop response should resemble a canonical HRF since each trial has a single stimulus and is short. The task timing and some sample trials are shown below (this is Figure 1 in a preprint (accepted version) which has much more task detail; see also the DMCC55B description).
So far, so good: the estimated responses for each task have a sensible shape, reflecting both the HRF and task timing. But what exactly is along the x-axis for the curves? And how did we fit these GLMs, and then decide to change them a bit? ... stay tuned.
later posts in this series: #2
UPDATE 4 January 2021: Corrected DMCC13benchmark openneuro links.
No comments:
Post a Comment