Real-time feedback reduces participant motion during task-based fMRI

The potential negative impact of head movement during fMRI has long been appreciated. Although a variety of prospective and retrospective approaches have been developed to help mitigate these effects, reducing head movement in the first place remains the most appealing strategy for optimizing data quality. Real-time interventions, in which participants are provided feedback regarding their scan-to-scan motion, have recently shown promise in reducing motion during resting state fMRI. However, whether feedback might similarly reduce motion during task-based fMRI is an open question. In particular, it is unclear whether participants can effectively monitor motion feedback while attending to task-related demands. Here we assessed whether a combination of real-time and between-run feedback could reduce head motion during task-based fMRI. During an auditory word repetition task, 78 adult participants (aged 19–81) were pseudorandomly assigned to receive feedback or not. Feedback was provided FIRMM software that used real-time calculation of realignment parameters to estimate participant motion. We quantified movement using framewise displacement (FD). We found that motion feedback resulted in a statistically significant reduction in participant head motion, with a small-to-moderate effect size (reducing average FD from 0.347 to 0.282). Reductions were most apparent in high-motion events. We conclude that under some circumstances real-time feedback may reduce head motion during task-based fMRI, although its effectiveness may depend on the specific participant population and task demands of a given study.


Introduction 67
Head movement during fMRI can introduce a number of different challenges to data analysis. 68 Although long appreciated (Friston et al., 1996), the detrimental influence of head motion on 69 resting state functional connectivity has only recently been widely recognized (Power et al., 70 2015). Rigid body realignment-a mainstay of fMRI analysis for decades-goes some way 71 towards improving correspondence across images (Ashburner and Friston, 2004), but does not 72 remove extraneous signal components introduced by movement (Friston et al., 1996). A 73 common approach for mitigating motion-related artifacts is to include the 6 realignment 74 parameters (translation and rotation around the X, Y, and Z axes) as nuisance regressors in 75 first-level models. Additional approaches include wavelet despiking (Patel et al., 2014), ICA 76 (Pruim et al., 2015), robust-weighted least squares (Diedrichsen and Shadmehr, 2005), 77 Bayesian approaches (Eklund et al., 2017), and frame censoring (Lemieux et al., 2007). 78 However, recent work investigating motion correction strategies in multiple data sets suggests 79 that the optimal strategy may depend on the specific data set and output metric (Jones et al.,80 2022), precluding a simple one-size-fits-all solution. 81 Given the lack of certainty regarding how to best handle head motion during analysis, 82 reducing head motion in the first place is a particularly appealing option. Common approaches 83 to reducing head motion during MRI scanning include foam padding and instructing participants 84 to not move during the scan. Other existing strategies include using custom head molds (Power 85 et al., 2019) and playing participants a relaxing abstract movie (Vanderwal et al., 2015). 86 A complementary approach for reducing head motion is to provide participants with 87 feedback regarding their head movement so they can learn to minimize motion. FIRMM 88 software 1 (Dosenbach et al., 2017) is one implementation of motion feedback. FIRMM uses 89 rapid image reconstruction and rigid-body alignment to estimate frame-by-frame movement, 90 providing visual cues to participants based on estimated movement. Feedback can either be 91 provided in real time, or between scanning runs. FIRMM has been shown to reduce movement 92 during resting state scans in both adults (Dosenbach et al., 2017) and in young children To evaluate the degree to which feedback might affect head motion during task-based 114 fMRI, we assigned participants in a task-based fMRI study of spoken word recognition to 115 receive feedback (the feedback group) or not (the no-feedback group). All other instructions, 116 scanning parameters, and task requirements were identical. We included both young and older 117 adults, and a task (word repetition) that we anticipated would typically require head movement. 118 Finally, we were able to capitalize on the fact that each subject has many observations 119 (hundreds of frames) by using linear mixed effects analysis. 120 The procedure is modeled after that in Rogers et al. (2020) and illustrated in Figure 1. 142 Participants performed a word repetition task in which on every trial they heard a spoken word, 143

Method
presented in stationary background noise (3 dB SNR), and repeated it back aloud. 144 The no-feedback group was given the following instructions: 145 146 During this task, it is important that you hold your body and head very still. 147 Please stay relaxed, stay alert, and keep your eyes open and on the fixation 148 cross. 149 150 151 Figure 1. Summary of task design and processing pipeline. a) Schematic of the word repetition task. Participants heard a word presented in the midst of background noise during the gap between volume acquisitions; during the following gap, they repeated the word aloud. b) Illustration of the two groups to which a participant could be pseudorandomly assigned. The group without feedback received standard instructions to keep still at the beginning of the study; the feedback group received instructions regarding real-time feedback, a colored crosshair indicating their scan-to-scan motion (framewise displacement; FD) during each run; and a visual report on feedback at the end of each run. c) Illustration of time variables available for modeling. 152 FIRMM was run in the background for experimenters to monitor motion during scanning 153 sequences and participants were given verbal feedback after each sequence. 154 The group receiving motion feedback was given the following instructions: 155 156 It is very important to remain still during your MRI so that we can obtain clear 157 images. Even very small movements that you are not aware of can affect the 158 (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted August 31, 2023. ; https://doi.org/10.1101/2023.01.12.523791 doi: bioRxiv preprint image quality. While performing the task, you will be receiving feedback 159 corresponding to your ability to remain still. This is to help you be aware of any 160 movements you may be making. You will see a white fixation cross on the 161 screen. The cross will change to yellow and then red depending on how much 162 you are moving. It will go back to white if you become still again. Sometimes 163 even if you are doing your best you will still see a red cross. This may mean that 164 the computer is too strict and not that you are necessarily doing anything wrong. 165 Just keep trying your best to keep the cross in the white. 166 167 The same researchers participated in data collection for all sessions. During the task, 168 participants viewed a colored cross in the middle of the screen. FD thresholds were set to show 169 a white cross at < 0.2 mm, a yellow cross at 0.2 mm to < 0.3 mm, and a red cross at ≥ 0.3 mm. 170 At the conclusion of each run, participants viewed a Head Motion Report (Supplemental Figure  171 1) which displayed their performance on a gauge of 0-100 (a percentage score of 0% to 100%), 172 and a graph of their motion level over time to help visualize total movement during the session. 173 Participants were encouraged to bring their score closer to 100% on subsequent sequences. 174  (Wellcome Trust  188 Centre for Human Neuroimaging) version 7487 (RRID:SCR_007037). Functional images 189 underwent rigid body realignment (i.e., "motion correction"). We used framewise displacement 190

MRI acquisition and analysis
(FD) to parsimoniously summarize frame-by-frame motion, calculated by combining the six head 191 motion estimates obtained from realignment, with a dimensional conversion of the three 192 rotations assuming the head is a 50 mm sphere (Power et al., 2012), to produce a summary 193 distance metric. We supplemented FD values with differential variance (DVARS), a measure of 194 variations in image intensity. DVARS was calculated as the root-mean-squared of the time 195 difference in the BOLD signal calculated across the entire brain, before realignment (Smyser et 196 al., 2011). Although FD and DVARS are highly correlated, they are not identical (Jones et al.,197 2022), and we anticipated that these two metrics might differ in sensitivity to motion feedback. 198

199
Movement parameters for young and older adults as a function of feedback are shown in Figure  200 2. Figure 2a shows FD values for two subjects (our first young adult without feedback, and our 201 first young adult with feedback), as well as the mean FD for all subjects in each feedback group. 202 Although we used all time points in the statistical analysis, for illustrative purposes we also 203 plotted summary values in Figure 2b (each point denoting mean FD per subject). 204 205 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted August 31, 2023. We analyzed these data using R version 4.2 (R Core Team, 2020) (RRID:SCR_001905). 211 We conducted linear mixed effects analysis using the nlme package (Pinheiro and Bates, 2000). 212 Because of the skewed distribution of FD data, we Gaussianized the FD values (Georg, 2011) 213 prior to modeling using the LambertW package (Georg, 2022) (see Supplemental Figure 1). 214 We accounted for temporal autocorrelation (that is, motion at one frame is related to motion 215 during the following frame) using a first-order autoregressive model.  (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made convenient to think about continuous scanning runs (sometimes also referred to as "sessions"). 236 In the current study, each participant had 6 runs of data. We thus ran an additional model to see 237 whether the effect of feedback was comparable across run. Because time was expressed as 6 238 runs, we did not account for temporal autocorrelation in this model: 239 240 m3 <-lme(FD ~ 1 + run + age_group * feedback, 241 random = ~ run | subject_number, 242 data = df, 243 method = "ML", 244 correlation = NULL) 245 246 247 Results of this analysis by run are shown in Table 3, and FD as a function of run are 248 shown in Figure 3. We again found a significant effect of feedback. 249 250 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made   Even though our analysis was focused on changes in mean FD, we wanted to look in 256 more detail at how feedback affected the distribution of movement. We thus plotted the density 257 of FD values for all scans as a function of feedback (Figure 4) for FD values up to 2.0. We used 258 density rather than the count to control for the different numbers of frames in the two groups 259 (resulting from different numbers of subjects). Although qualitative, this analysis highlights that 260 participants receiving feedback showed fewer high-FD scans than those receiving lower 261 feedback, with this effect becoming more apparent at higher FD values-for FD values above 262 1.5, the feedback group had less than a third of these scans than the group without feedback. 263 264 265 We next repeated our primary FD analyses, but using DVARS (rather than FD) as a 268 dependent measure, under the assumption that FD and DVARS may be differently sensitive to 269 effects of motion, and one or the other may be of more interest in any specific study. As with 270 FD, we used Gaussianized data to reduce the influence of the skewed distribution. Summary 271 results are shown in Figure 5 and Table 4. Results as a function of run are additionally shown 272 in Figure 6 and Table 5. Although DVARS was numerically lower for the feedback group than 273 the no-feedback group, the difference was not statistically significant.    In addition to framewise measurements of FD and DVARS, we investigated whether 282 motion feedback affected the temporal signal-to-noise ratio (tSNR) of the data. We first created 283 a brain mask using each subject's tissue class segmentation, binarizing at a gray matter 284 probability of 0.8. We then calculated tSNR (the mean of the signal divided by its standard 285 deviation, over frames) in every voxel within the brain mask. We then performed two 286 complementary analyses. First, for each run, we took the mean tSNR over voxels, and used the 287 same statistical model as for FD and DVARS, shown in Figure 7 and shown in Table 6. 288 Although the tSNR values were numerically higher in the feedback group, there was no 289 significant effect of feedback (p = 0.0814). (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted August 31, 2023. ; https://doi.org/10.1101/2023.01.12.523791 doi: bioRxiv preprint the behavioral task. Accuracy on the word repetition task is shown in Figure 8 as a function of 302 age group and feedback. We used a linear model to test for the impact of feedback on 303 behavioral accuracy (averaging over experimental runs). We found no significant effect of age 304 group, no significant effect of feedback, and no significant age x feedback interaction, although 305 in older adults there was a small numerical decrease in accuracy during the feedback condition. 306 307  (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted August 31, 2023. ; https://doi.org/10.1101/2023.01.12.523791 doi: bioRxiv preprint

Discussion 309
Given the pernicious and sometimes unpredictable effects of head movement in fMRI studies, 310 reducing motion in the first place is on its face an appealing strategy. In the current study we 311 evaluated the degree to which a combination of real-time and summary motion feedback could 312 reduce head movement in adult participants by pseudorandomly assigning participants to 313 feedback or no feedback groups. We found that motion feedback was associated with a 314 statistically significant reduction in average FD of approximately 0.06, with consistent (though 315 not significant) changes in DVARS and tSNR. We discuss these findings and their implications 316 below. 317 First, our primary finding that providing motion feedback can significantly reduce head 318 motion in participants undergoing task-based MRI is notable. Our findings are also consistent 319 with those of Yang and colleagues (2005), who reported a reduction in head motion after 320 integrating visual feedback in the display of an N-back task. That being said, the principal effect of feedback may not be the average reduction in FD 334 observed, but rather a reduction in the number of large motion events. Rapid head motion tends 335 to have a larger effect on the BOLD signal than slower motion. Thus, reducing the number of 336 large, quick movements may improve signal quality to a greater extent than a simple reduction 337 in mean FD. Indeed, we saw the biggest proportional difference between feedback groups at 338 higher FD values (Figure 4b). Interestingly, the reduction in high motion events roughly 339 corresponds to an FD of 0.3, the value at which participants received "red" feedback. 340 An important consideration when implementing real-time feedback during fMRI is 341 whether doing so changes task demands. Although we found accuracy was statistically 342 comparable across feedback groups, we cannot rule out other changes. Indeed, we observed a 343 numeric (albeit not significant) decrease in task accuracy in older adults when they were 344 presented with real-time feedback relative to the no-feedback group. Broadly speaking, we 345 might expect monitoring the display for feedback and adjusting one's movement in response to 346 engage systems related to cognitive control (sometimes also referred to as attentional control or 347 executive function) (see also McCabe et al., 2010). In one framing of cognitive control, cognitive 348 control is proposed to operate in at least two complementary modes: a proactive mode that is 349 concerned with maintaining goal-directed behavior, and a reactive mode that is engaged when 350 goal-directed performance falters (Braver, 2012). There are two implications of such a 351 framework. First, participants who are less able to engage cognitive control may benefit less 352 from motion feedback compared to those who are better able to engage control systems. 353 Second, researcher concerns about how task demands may affect imaging results of interest 354 may depend on the nature of the main fMRI task. For example, tasks tapping cognitive control 355 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted August 31, 2023. ; https://doi.org/10.1101/2023.01.12.523791 doi: bioRxiv preprint or executive function may be more impacted by motion feedback than those focused on primary 356 sensory or motor systems. In our case, we used a speech-in-noise task, one that is 357 hypothesized to engage cognitive control networks (Peelle, 2018;McLaughlin et al., 2021). 358 We did notice a trend towards age effects, such that older adults' behavioral 359 performance was more affected than that of young adults in the feedback condition. In this 360 context it may also be relevant that older adults can be more sensitive to errors (and thus 361 differentially sensitive to feedback). For example, in drift diffusion models of cognitive tasks, 362 older adults often differ in threshold than drift rate compared to young adults (e.g., Ratcliff et al.,363 2004). 364 In our implementation, participants received brief instructions before entering the 365 scanner, and once in the scanner the display was explained to them. One interesting future 366 direction would be to provide participants with motion training over a longer period to see 367 whether they might learn to better limit their motion in the absence of feedback (for example, by 368 training in a mock scanner). Such an approach would take more time, but might circumvent 369 some of the challenges associated with real-time feedback (such as introducing a "dual-task" 370 situation) (Pashler, 1994 is consistent with the effect we saw for FD. In addition, we found that feedback led to increased 377 whole-brain tSNR, which we used as a proxy of overall image quality. Despite the tSNR being 378 numerically larger in the group receiving motion feedback, the difference was not statistically 379 significant. However, it is still possible that small improvements in tSNR would result in 380 appreciable improvements in the accuracy or reliability of statistical models. 381 There are several important caveats associated with the current report. One is that the 382 instructions were not perfectly matched across group; specifically, the group receiving feedback 383 was told that "[e]ven very small movements that you are not aware of can affect the image 384 quality", an instruction not given to the group not receiving feedback. It could be that the specific 385 wording of the instructions affected participant behavior. A related point pertains to our task 386 design. We opted for a between-subjects design to avoid any "contamination" of motion 387 feedback. However, this choice necessarily resulted in groups that contained different 388 individuals (and in our case, different instructions). An alternative design would be a within-389 subjects manipulation, which may be a fruitful area for further extensions of this work. 390 Finally, it is worth considering real-time feedback using estimated motion, as 391 implemented here, compares with other real-time feedback approaches. One particularly 392 attractive approach was introduced by Krause and colleagues (2019), in which the authors 393 affixed medical tape to participants' heads. The adhesion of the tape provided tactile feedback 394 when participants moved their heads, and this real-time feedback significantly reduced 395 participant's head motion (both translation and rotation). An added advantage of this approach 396 is the simplicity of implementation and lack of custom software (e.g., real-time analysis of 397 movement parameters) or interference with task (i.e., tactile feedback does not interfere with 398 visual stimuli). Individual researchers will need to decide which approach to motion reduction, if 399 any, is appropriate for a given task and specific population. 400 In conclusion, we found that motion feedback as implemented in FIRMM significantly 401 reduced the amount of head motion we observed during task-based fMRI. Real-time feedback 402 may thus be well-suited to complement other approaches to motion reduction depending on the 403 specific needs of a given study. 404 . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted August 31, 2023. ; https://doi.org/10.1101/2023.01.12.523791 doi: bioRxiv preprint