Neuroconscience

The latest thoughts, musings, and data in cognitive science and neuroscience.

When is expectation not a confound? On the necessity of active controls.

Learning and plasticity are hot topics in neuroscience. Whether exploring old world wisdom or new age science fiction, the possibility that playing videogames might turn us into attention superheroes or that practicing esoteric meditation techniques might heal troubled minds is an exciting avenue for research. Indeed findings suggesting that exotic behaviors or novel therapeutic treatments might radically alter our brain (and behavior) are ripe for sensational science-fiction headlines purporting vast brain benefits.  For those of you not totally bored of methodological crisis, here we have one brewing anew. You see the standard recommendation for those interested in intervention research is the active-controlled experimental design. Unfortunately in both clinical research on psychotherapy (including meditation) and more Sci-Fi areas of brain training and gaming, use of active controls is rare at best when compared to the more convenient (but causally ineffective) passive control group. Now a new article in Perspectives in Psychological Science suggests that even standard active controls may not be sufficient to rule out confounds in the treatment effect of interest.

Why is that? And why exactly do we need  active controls in the first place? As the authors clearly point out, what you want to show with such a study is the causal efficacy of the treatment of interest. Quite simply what that means is that the thing you think should have some interesting effect should actually be causally responsible for creating that effect. If you want to argue that standing upside down for twenty minutes a day will make me better at playing videogames in Australia, it must be shown that it is actually standing upside down that causes my increased performance down under. If my improved performance on Minecraft Australian Edition is simply a product of my belief in the power of standing upside down, or my expectation that standing upside down is a great way to best kangaroo-creepers, then we have no way of determining what actually produced that performance benefit. Research on placebos and the power of expectations shows that these kinds of subjective beliefs can have a big impact on everything from attentional performance to mortality rates.

Useful flowchart from Boot et al on whether or not a study can make causal claims for treatment.

Useful flowchart from Boot et al on whether or not a study can make causal claims for treatment.

Typically researchers attempt to control for such confounds through the use of a control group performing a task as similar as possible to the intervention of interest. But how do we know participants in the two groups don’t end up with different expectations about how they should improve as a result of the training? Boot et al point out that without actually measuring these variables, we have no idea and no way of knowing for sure that expectation biases don’t produce our observed improvements. They then provide a rather clever demonstration of their concern, in an experiment where participants view videos of various cognition tests as well as videos of a training task they might later receive, in this case either the first-person shooter Unreal Tournament or the spatial puzzle game Tetris. Finally they asked the participants in each group which tests they thought they’d do better on as a result of the training video. Importantly the authors show that not only did UT and Tetris lead to significantly different expectations, but also that those expectation benefits were specific to the modality of trained and tested tasks. Thus participant who watched the action-intensive Unreal Tournament videos expected greater improvements on tests of reaction time and visual performance, whereas participants viewing Tetris rated themselves as likely to do better on tests of spatial memory.

This is a critically important finding for intervention research. Many researchers, myself included, have often thought of the expectation and demand characteristic confounds in a rather general way. Generally speaking until recently I wouldn’t have expected the expectation bias to go much beyond a general “I’m doing something effective” belief. Boot et al show that our participants are a good deal cleverer than that, forming expectations-for-improvement that map onto specific dimensions of training. This means that to the degree that an experimenter’s hypothesis can be discerned from either the training or the test, participants are likely to form unbalanced expectations.

The good news is that the authors provide several reasonable fixes for this dilemma. The first is just to actually measure participant’s expectations, specifically in relation to the measures of interest. Another useful suggestion is to run pilot studies ensuring that the two treatments do not evoke differential expectations, or similarly to check that your outcome measures are not subject to these biases. Boot and colleagues throw the proverbial glove down, daring readers to attempt experiments where the “control condition” actually elicits greater expectations yet the treatment effect is preserved. Further common concerns, such as worries about balancing false positives against false negatives, are address at length.

The entire article is a great read, timely and full of excellent suggestions for caution in future research. It also brought something I’ve been chewing on for some time quite clearly into focus. From the general perspective of learning and plasticity, I have to ask at what point is an expectation no longer a confound. Boot et al give an interesting discussion on this point, in which they suggest that even in the case of balanced expectations and positive treatment effects, an expectation dependent response (in which outcome correlates with expectation) may still give cause for concern as to the causal efficacy of the trained task. This is a difficult question that I believe ventures far into the territory of what exactly constitutes the minimal necessary features for learning. As the authors point out, placebo and expectations effects are “real” products of the brain, with serious consequences for behavior and treatment outcome. Yet even in the medical community there is a growing understanding that such effects may be essential parts of the causal machinery of healing.

Possible outcome of a training experiment, in which the control shows no dependence between expectation and outcome (top panel) and the treatment of interest shows dependence (bottom panel). Boot et al suggest that such a case may invalidate causal claims for treatment efficacy.

Possible outcome of a training experiment, in which the control shows no dependence between expectation and outcome (top panel) and the treatment of interest shows dependence (bottom panel). Boot et al suggest that such a case may invalidate causal claims for treatment efficacy.

To what extent might this also be true of learning or cognitive training? For sure we can assume that expectations shape training outcomes, otherwise the whole point about active controls would be moot. But can one really have meaningful learning if there is no expectation to improve? I realize that from an experimental/clinical perspective, the question is not “is expectation important for this outcome” but “can we observe a treatment outcome when expectations are balanced”. Still when we begin to argue that the observation of expectation-dependent responses in a balanced design might invalidate our outcome findings, I have to wonder if we are at risk of valuing methodology over phenomena. If expectation is a powerful, potentially central mechanism in the causal apparatus of learning and plasticity, we shouldn’t be surprised when even efficacious treatments are modulated by such beliefs. In the end I am left wondering if this is simply an inherent limitation in our attempt to apply the reductive apparatus of science to increasingly holistic domains.

Please do read the paper, as it is an excellent treatment of a critically ignored issue in the cognitive and clinical sciences. Anyone undertaking related work should expect this reference to appear in reviewer’s replies in the near future.

EDIT:
Professor Simons, a co-author of the paper, was nice enough to answer my question on twitter. Simons pointed out that a study that balanced expectation, found group outcome differences, and further found correlations of those differences with expectation could conclude that the treatment was causally efficacious, but that it also depends on expectations (effect + expectation). This would obviously be superior to an unbalanced designed or one without measurement of expectation, as it would actually tell us something about the importance of expectation in producing the causal outcome. Be sure to read through the very helpful FAQ they’ve posted as an addendum to the paper, which covers these questions and more in greater detail. Here is the answer to my specific question:

What if expectations are necessary for a treatment to work? Wouldn’t controlling for them eliminate the treatment effect?

No. We are not suggesting that expectations for improvement must be eliminated entirely. Rather, we are arguing for the need to equate such expectations across conditions. Expectations can still affect the treatment condition in a double-blind, placebo-controlled design. And, it is possible that some treatments will only have an effect when they interact with expectations. But, the key to that design is that the expectations are equated across the treatment and control conditions. If the treatment group outperforms the control group, and expectations are equated, then something about the treatment must have contributed to the improvement. The improvement could have resulted from the critical ingredients of the treatment alone or from some interaction between the treatment and expectations. It would be possible to isolate the treatment effect by eliminating expectations, but that is not essential in order to claim that the treatment had an effect.

In a typical psychology intervention, expectations are not equated between the treatment and control condition. If the treatment group improves more than the control group, we have no conclusive evidence that the ingredients of the treatment mattered. The improvement could have resulted from the treatment ingredients alone, from expectations alone, or from an interaction between the two. The results of any intervention that does not equate expectations across the treatment and control condition cannot provide conclusive evidence that the treatment was necessary for the improvement. It could be due to the difference in expectations alone. That is why double blind designs are ideal, and it is why psychology interventions must take steps to address the shortcomings that result from the impossibility of using a double blind design. It is possible to control for expectation differences without eliminating expectations altogether.

Can the reward prediction error hypothesis explain addiction and reward? Great video on incentive-salience theory.

If you are interested in predictive coding, learning, motivation, addiction, or reward, don’t miss this excellent video by Kent Berridge. The incentive salience theory has long fascinated me as it may potentially explain data not accounted for by the hedonic-aversive accounts of addiction and reward. Essentially Incentive Salience argues that rather than reward or addiction being purely the function of seeking hedonic rewards or avoiding aversive punishments (e.g. withdrawal), salient cues trigger direct “wanting” responses irrespective of the reward itself. This may explain why for example, mice often continue to seek rewards even when dopamine is blocked and no pleasurable outcome can be obtained, or conversely why even in the absence of withdrawal addicts will seek drugs in response to salient cues.  Anyway, Kent explains it much better than I can- so watch it!

Can compassion be trained like a muscle? Active-controlled fMRI of compassion meditation.

Among the cognitive training literature, meditation interventions are particularly unique in that they often emphasize emotional or affective processing at least as much as classical ‘top-down’ attentional control. From a clinical and societal perspective, the idea that we might be able to “train” our “emotion muscle” is an attractive one. Recently much has been made of the “empathy deficit” in the US, ranging from empirical studies suggesting a relationship between quality-of-care and declining caregiver empathy, to a recent push by President Obama to emphasize the deficit in numerous speeches.

While much of the training literature focuses on cognitive abilities like sustained attention and working memory, many investigating meditation training have begun to study the plasticity of affective function, myself included.  A recent study by Helen Weng and colleagues in Wisconsin investigated just this question, asking if compassion (“loving-kindness”) meditation can alter altruistic behavior and associated neural processing. Her study is one of the first of its kind, in that rather than merely comparing groups of advanced practitioners and controls, she utilized a fully-randomized active-controlled design to see if compassion responds to brief training in novices while controlling for important confounds.

As many readers should be aware, a chronic problem in training studies is a lack of properly controlled longitudinal design. At best, many rely on “passive” or “no-contact” controls who merely complete both measurements without receiving any training. Even in the best of circumstances “active” controls are often poorly matched to whatever is being emphasized and tested in the intervention of interest. While having both groups do “something” is better than a passive or no-control design, problems may still arise if the measure of interest is mismatched to the demand characteristics of the study.  Stated simply, if your condition of interest receives attention training and attention tests, and your control condition receives dieting instruction or relaxation, you can expect group differences to be confounded by an explicit “expectation to improve” in the interest group.

In this regard Weng et al present an almost perfect example of everything a training study should be. Both interventions were delivered via professionally made audio CDs (you can download them yourselves here!), with participants’ daily practice experiences being recorded online. The training materials were remarkably well matched for the tests of interest and extra care was taken to ensure that the primary measures were not presented in a biased way. The only thing they could have done further would be a single blind (making sure the experimenters didn’t know the group identity of each participant), but given the high level of difficulty in blinding these kinds of studies I don’t blame them for not undertaking such a manipulation. In all the study is extremely well-controlled for research in this area and I recommend it as a guideline for best practices in training research.

Specifically, Weng et al tested the impact of loving-kindness compassion meditation or emotion reappraisal training on an emotion regulation fMRI task and behavioral economic game measuring altruistic behavior. For the fMRI task, participants viewed emotional pictures (IAPS) depicting suffering or neutral scenarios and either practiced a compassion meditation or reappraisal strategy to regulate their emotional response, before and after training. After the follow-up scan, good-old fashion experimental deception was used to administer a dictator economics-game that was ostensibly not part of the primary study and involved real live players (both deceptions).

For those not familiar with the dictator game, the concept is essentially that a participant watches a “dictator” endowed with 100$ give “unfair” offers to a “victim” without any money. Weng et al took great care in contextualizing the test purely in economic terms, limiting demand confounds:

Participants were told that they were playing the game with live players over the Internet. Effects of demand characteristics on behavior were minimized by presenting the game as a unique study, describing it in purely economic terms, never instructing participants to use the training they received, removing the physical presence of players and experimenters during game play, and enforcing real monetary consequences for participants’ behavior.

This is particularly important, as without these simple manipulations it would be easy for stodgy reviewers like myself to worry about subtle biases influencing behavior on the task. Equally important is the content of the two training programs. If for example, Weng et al used a memory training or attention task as their active-control group, it would be difficult not to worry that behavioral differences were due to one group expecting a more emotional consequence of the study, and hence acting more altruistic. In the supplementary information, Weng et al describe the two training protocols in great detail:

Compassion

… Participants practiced compassion for targets by 1) contemplating and envisioning their suffering and then 2) wishing them freedom from that suffering. They first practiced compassion for a Loved One, such as a friend or family member. They imagined a time their loved one had suffered (e.g., illness, injury, relationship problem), and were instructed to pay attention to the emotions and sensations this evoked. They practiced wishing that the suffering were relieved and repeated the phrases, “May you be free from this suffering. May you have joy and happiness.” They also envisioned a golden light that extended from their heart to the loved one, which helped to ease his/her suffering. They were also instructed to pay attention to bodily sensations, particularly around the heart. They repeated this procedure for the Self, a Stranger, and a Difficult Person. The Stranger was someone encountered in daily life but not well known (e.g., a bus driver or someone on the street), and the Difficult Person was someone with whom there was conflict (e.g., coworker, significant other). Participants envisioned hypothetical situations of suffering for the stranger and difficult person (if needed) such as having an illness or experiencing a failure. At the end of the meditation, compassion was extended towards all beings. For each new meditation session, participants could choose to use either the same or different people for each target category (e.g., for the loved one category, use sister one day and use father the next day).

Reappraisal

… Participants were asked to recall a stressful experience from the past 2 years that remained upsetting to them, such as arguing with a significant other or receiving a lower-than- expected grade. They were instructed to vividly recall details of the experience (location, images, sounds). They wrote a brief description of the event, and chose one word to best describe the feeling experienced during the event (e.g., sad, angry, anxious). They rated the intensity of the feeling during the event, and the intensity of the current feeling on a scale (0 = No feeling at all, 100 = Most intense feeling in your life). They wrote down the thoughts they had during the event in detail. Then they were asked to reappraise the event (to think about it in a different, less upsetting way) using 3 different strategies, and to write down the new thoughts. The strategies included 1) thinking about the situation from another person’s perspective (e.g., friend, parent), 2) viewing it in a way where they would respond with very little emotion, and 3) imagining how they would view the situation if a year had passed, and they were doing very well. After practicing each strategy, they rated how reasonable each interpretation was (0 = Not at all reasonable, 100 = Completely reasonable), and how badly they felt after considering this view (0 = Not bad at all, 100 = Most intense ever). Day to day, participants were allowed to practice reappraisal with the same stressful event, or choose a different event. Participants logged the amount of minutes practiced after the session.

In my view the active control is extremely well designed for the fMRI and economic tasks, with both training methods explicitly focusing on the participant altering an emotional response to other individuals.  In tests of self-rated efficacy, both groups showed significant decreases in negative emotion, further confirming the active control. Interestingly when Weng et al compared self-ratings over time, only the compassion group showed significant reduction from the first half of training sessions to the last. I’m not sure if this constitutes a limitation, as Weng et al further report that on each individual training day the reappraisal group reported significant reductions, but that the reductions themselves did not differ significantly over time. They explain this as being likely due to the fact that the reappraisal group frequently changed emotional targets, whereas the compassion group had the same 3 targets throughout the training. Either way the important point is that both groups self-reported similar overall reductions in negative emotion during the course of the study, strongly supporting the active control.

Now what about the findings? As mentioned above, Weng et al tested participants before and after training on an fMRI emotion regulation task. After the training, all participants performed the “dictator game”, shown below. After rank-ordering the data, they found that the compassion group showed significantly greater redistribution:

The dictator task (left) and increased redistribution (right).

For the fMRI analysis, they analyzed BOLD responses to negative vs neutral images at both time points, subtracted the beta coefficients, and then entered these images into a second-level design matrix testing the group difference, with the rank-ordered redistribution scores as a covariate of interest. They then tested for areas showing group differences in the correlation of redistribution scores and changes of BOLD response to negative vs neutral images (pre vs post), across the whole brain and in several ROIs, while properly correcting for multiple comparisons. Essentially this analysis asks, where in the brain do task-related changes in BOLD correlate more or less with the redistribution score in one group or another. For the group x covariate interaction they found significant differences (increased BOLD-covariate correlation) in the right inferior parietal cortex (IPC), a region of the parietal attention network, shown on the left-hand panel:

Increased correlation between negative vs neutral imagery related BOLD and redistribution scores (left), connectivity with DLPFC (right).

They further extracted signal from the IPC cluster and entered it into a conjunction analysis, testing for areas showing significant correlation  with the IPC activity, and found a strong effect in right DLPFC (right panel). Finally they performed a psychophysiological interaction (PPI) analysis with the right DLPFC activity as the seed, to determine regions showing significant task-modulated connectivity with that DLPFC activity. The found increased emotion-modulated DLPFC connectivity to nucleus accumbens, a region involved in encoding positive rewards (below, right).

Screen shot 2013-05-23 at 3.21.15 PM

PPI shows increased emotion-modulated connectivity of nucleus accumbens and DLPFC.

Together these results implicate training-related BOLD activity increases to emotional stimuli in the parietal attention network and increased parietal connectivity with regions implicated in cognitive control and reward processing, in the observed altruistic behavior differences. The authors conclude that compassion training may alter emotional processing through a novel mechanism, where top-down central-executive circuits redirect emotional information to areas associated with positive reward, reflecting the role of compassion meditation in emphasizing increased positive emotion to the aversive states of others. A fitting and interesting conclusion, I think.

Overall, the study should receive high marks for its excellent design and appropriate statistical rigor. There is quite a bit of interesting material in the supplementary info, a strategy I dislike, but that is no fault of the authors considering the publishing journal (Psych Science). The question itself is extremely novel, in terms of previous active-controlled studies. To date only one previous active-controlled study investigated the role of compassion meditation on empathy-related neuroplasticity. However that study compared compassion meditation with a memory strategy course, which (in my opinion) exposes it to serious criticism regarding demand characteristic. The authors do reference that study, but only briefly to state that both studies support a role of compassion training in altering positive emotion- personally I would have appreciated a more thorough comparison, though I suppose I can go and to that myself if I feel so inclined :).

The study does have a few limitations worth mentioning. One thing that stood out to me was that the authors never report the results of the overall group mean contrast for negative vs neutral images. I would have liked to know if the regions showing increased correlation with redistribution actually showed higher overall mean activation increases during emotion regulation. However as the authors clearly had quite specific hypotheses, leading them to  restrict their alpha to 0.01 (due to testing 1 whole-brain contrast and 4 ROIs), I can see why they left this out. Given the strong results of the study, it would in retrospect perhaps have been more prudent to skip  the ROI analysis (which didn’t seem to find much) and instead focus on testing the whole brain results.  I can’t blame them however, as it is surprising not to see anything going on in insula or amygdala for this kind of training.  It is also a bit unclear to me why the DLPFC was used as the PPI seed as opposed to the primary IPC cluster, although I am somewhat unfamiliar with the conjunction-connectivity analysis used here. Finally, as the authors themselves point out, a major limitation of the study is that the redistribution measure was collected only at time two, preventing a comparison to baseline for this measure.

Given the methodological state of the topic (quite poor, generally speaking), I am willing to grant them these mostly minor caveats. Of course, without a baseline altruism measure it is difficult to make a strong conclusion about the causal impact of the meditation training on altruism behavior, but at least their neural data are shielded from this concern. So while we can’t exhaustively conclude that compassion can be trained, the results of this study certainly suggest it is possible and perhaps even likely, providing a great starting point for future research. One interesting thing for me was the difference in DLPFC. We also found task-related increases in dorsolateral prefrontal cortex following active-controlled meditation, although in the left hemisphere and for a very different kind of training and task. One other recent study of smoking cessation also reported alteration in DLPFC following mindfulness training, leading me to wonder if we’re seeing the emergence of empirical consensus for this region’s specific involvement in meditation training. Another interesting point for me was that affective regulation here seems to involve primarily top-down or attention related neural correlates,  suggesting that bottom-up processing (insula, amygdala) may be more resilient to brief training, something we also found in our study. I wonder if the group mean-contrasts would have been revealing here (i.e. if there were differences in bottom-up processing that don’t correlate with redistribution). All together a great study that raises the bar for training research in cognitive neuroscience!

Is the resting BOLD signal physiological noise? What about resting EEG?

Over the past 5 years, resting-state fMRI (rsfMRI) has exploded in popularity. Literally dozens of papers are published each day examining slow (< .1 hz) or “low frequency” fluctuations in the BOLD signal. When I first moved to Europe I was caught up in the somewhat North American frenzy of resting state networks. I couldn’t understand why my Danish colleagues, who specialize in modelling physiological noise in fMRI, simply did not take the literature seriously. The problem is essentially that the low frequencies examined in these studies are the same as those that dominate physiological rhythms. Respiration and cardiac pulsation can make up a massive amount of variability in the BOLD signal. Before resting state fMRI came along, nearly every fMRI study discarded any data frequencies lower than one oscillation every 120 seconds (e.g. 1/120 Hz high pass filtering). Simple things like breath holding and pulsatile motion in vasculature can cause huge effects in BOLD data, and it just so happens that these artifacts (which are non-neural in origin) tend to pool around some of our favorite “default” areas: medial prefrontal cortex, insula, and other large gyri near draining veins.

Naturally this leads us to ask if the “resting state networks” (RSNs) observed in such studies are actually neural in origin, or if they are simply the result of variations in breath pattern or the like. Obviously we can’t answer this question with fMRI alone. We can apply something like independent component analysis (ICA) and hope that it removes most of the noise- but we’ll never really be 100% sure we’ve gotten it all that way. We can measure the noise directly (e.g. “nuisance covariance regression”) and include it in our GLM- but much of the noise is likely to be highly correlated with the signal we want to observe. What we need are cross-modality validations that low-frequency oscillations do exist, that they drive observed BOLD fluctuations, and that these relationships hold even when controlling for non-neural signals. Some of this is already established- for example direct intracranial recordings do find slow oscillations in animal models. In MEG and EEG, it is well established that slow fluctuations exist and have a functional role.

So far so good. But what about in fMRI? Can we measure meaningful signal while controlling for these factors? This is currently a topic of intense research interest. Marcus Raichle, the ‘father’ of the default mode network, highlights fascinating multi-modal work from a Finnish group showing that slow fluctuations in behavior and EEG signal coincide (Raichle and Snyder 2007; Monto, Palva et al. 2008). However, we should still be cautious- I recently spoke to a post-doc from the Helsinki group about the original paper, and he stressed that slow EEG is just as contaminated by physiological artifacts as fMRI. Except that the problem is even worse, because in EEG the artifacts may be several orders of magnitude larger than the signal of interest[i].

Understandably I was interested to see a paper entitled “Correlated slow fluctuations in respiration, EEG, and BOLD fMRI” appear in Neuroimage today (Yuan, Zotev et al. 2013). The authors simultaneously collected EEG, respiration, pulse, and resting fMRI data in 9 subjects, and then perform cross-correlation and GLM analyses on the relationship of these variables, during both eyes closed and eyes open rest. They calculate Respiratory Volume per Time (RVT), a measure developed by Rasmus Birn, to assign a respiratory phase to each TR (Birn, Diamond et al. 2006). One key finding is that the global variations in EEG power are strongly predicted by RVT during eyes closed rest, with a maximum peak correlation coefficient of .40. Here are the two time series:

RVTalpha 

You can clearly see that there is a strong relationship between global alpha (GFP) and respiration (RVT). The authors state that “GFP appears to lead RVT” though I am not so sure. Regardless, there is a clear relationship between eyes closed ‘alpha’ and respiration. Interestingly they find that correlations between RVT and GFP with eyes open were not significantly different from chance, and that pulse did not correlate with GFP. They then conduct GLM analyses with RVT and GFP as BOLD regressors. Here is what their example subject looked like during eyes-closed rest:

RVT_GFP_BOLD

Notice any familiar “RSNs” in the RVT map? I see anti-correlated executive deactivation and default mode activation! Very canonical.  Too bad they are breath related. This is why noise regression experts tend to dislike rsfMRI, particularly when you don’t measure the noise. We also shouldn’t be too surprised that the GFP-BOLD and RVT-BOLD maps look similar, considering that GFP and RVT are highly correlated. After looking at these correlations separately, Yuan et al perform RETROICOR physiological noise correction and then reexamine the contrasts. Here are the group maps:

group_map

Things look a bit less default-mode-like in the group RVT map, but the RVT and GFP maps are still clearly quite similar. In panel D you can see that physiological noise correction has a large global impact on GFP-BOLD correlations, suggesting that quite a bit of this co-variance is driven by physiological noise. Put simply, respiration is explaining a large degree of alpha-BOLD correlation; any experiment not modelling this covariance is likely to produce strongly contaminated results. Yuan et al go on to examine eyes-open rest and show that, similar to their RVT-GFP cross-correlation analysis, not nearly as much seems to be happening in eyes open compared to closed:

eyesopen

The authors conclude that “In particular, this correlation between alpha EEG and respiration is much stronger in eyes-closed resting than in eyes-open resting” and that “[the] results also suggest that eyes-open resting may be a more favorable condition to conduct brain resting state fMRI and for functional connectivity analysis because of the suppressed correlation between low-frequency respiratory fluctuation and global alpha EEG power, therefore the low-frequency physiological noise predominantly of non-neuronal origin can be more safely removed.” Fair enough- one conclusion is certainly that eyes closed rest seems much more correlated with respiration than eyes open. This is a decent and useful result of the study. But then they go on to make this really strange statement, which appears in the abstract, introduction, and discussion:

“In addition, similar spatial patterns were observed between the correlation maps of BOLD with global alpha EEG power and respiration. Removal of respiration related physiological noise in the BOLD signal reduces the correlation between alpha EEG power and spontaneous BOLD signals measured at eyes-closed resting. These results suggest a mutual link of neuronal origin between the alpha EEG power, respiration, and BOLD signals”’ (emphasis added)

That’s one way to put it! The logic here is that since alpha = neural activity, and respiration correlates with alpha, then alpha must be the neural correlate of respiration. I’m sorry guys, you did a decent experiment, but I’m afraid you’ve gotten this one wrong. There is absolutely nothing that implies alpha power cannot also be contaminated by respiration-related physiological noise. In fact it is exactly the opposite- in the low frequencies observed by Yuan et al the EEG data is particularly likely to be contaminated by physiological artifacts! And that is precisely what the paper shows – in the author’s own words: “impressively strong correlations between global alpha and respiration”. This is further corroborated by the strong similarity between the RVT-BOLD and alpha-BOLD maps, and the fact that removing respiratory and pulse variance drastically alters the alpha-BOLD correlations!

So what should we take away from this study? It is of course inconclusive- there are several aspects of the methodology that are puzzling to me, and sadly the study is rather under-powered at n = 9. I found it quite curious that in each of the BOLD-alpha maps there seemed to be a significant artifact in the lateral and posterior ventricles, even after physiological noise correction (check out figure 2b, an almost perfect ventricle map). If their global alpha signal is specific to a neural origin, why does this artifact remain even after physiological noise correction? I can’t quite put my finger on it, but it seems likely to me that some source of noise remained even after correction- perhaps a reader with more experience in EEG-fMRI methods can comment. For one thing their EEG motion correction seems a bit suspect, as they simply drop outlier timepoints. One way or another, I believe we should take one clear message away from this study – low frequency signals are not easily untangled from physiological noise, even in electrophysiology. This isn’t a damnation of all resting state research- rather it is a clear sign that we need be to measuring these signals to retain a degree of control over our data, particularly when we have the least control at all.

References:

Birn, R. M., J. B. Diamond, et al. (2006). “Separating respiratory-variation-related fluctuations from neuronal-activity-related fluctuations in fMRI.” Neuroimage 31(4): 1536-1548.

Monto, S., S. Palva, et al. (2008). “Very slow EEG fluctuations predict the dynamics of stimulus detection and oscillation amplitudes in humans.” The Journal of Neuroscience 28(33): 8268-8272.

Raichle, M. E. and A. Z. Snyder (2007). “A default mode of brain function: a brief history of an evolving idea.” Neuroimage 37(4): 1083-1090.

Yuan, H., V. Zotev, et al. (2013). “Correlated Slow Fluctuations in Respiration, EEG, and BOLD fMRI.” NeuroImage pp. 1053-8119.

 


[i] Note that this is not meant to be in anyway a comprehensive review. A quick literature search suggests that there are quite a few recent papers on resting BOLD EEG. I recall a well done paper by a group at the Max Planck Institute that did include noise regressors, and found unique slow BOLD-EEG relations. I cannot seem to find it at the moment however!

 

MOOC on non-linear approaches to social and cognitive sciences. Votes needed!

My colleagues at Aarhus University have put together a fascinating proposal for a Massive Online Open Course (MOOC) on “Analyzing Behavioral Dynamics: non-linear approaches to social and cognitive sciences”. I’ve worked with Riccardo and Kristian since my masters and I can promise you the course will be excellent. They’ve spent the past 5 years exhaustively pursuing methodology in non-linear dynamics, graph theoretical, and semantic/semiotic analyses and I think will have a lot of interesting practical insights to offer. Best of all the course is free to all, as long as it gets enough votes on the MPF website. I’ve been a bit on the fence regarding my feelings about MOOCs, but in this case I think it’s really a great opportunity to give novel methodologies more exposure. Check it out- if you like it, give them a vote and consider joining the course!

https://moocfellowship.org/submissions/analyzing-behavioral-dynamics-non-linear-approaches-to-social-and-cognitive-sciences

Course Description

In the last decades, the social sciences have come to confront the temporal nature of human behavior and cognition: How do changes of heartbeat underlie emotions? How do we regulate our voices in a conversation? How do groups develop coordinative strategies to solve complex problems together?
This course enables you to tackle this sort of questions: addresses methods of analysis from nonlinear dynamics and complexity theory, which are designed to find and characterize patterns in this kind of complicated data. Traditionally developed in fields like physics and biology, non-linear methods are often neglected in social and cognitive sciences.

The course consists of two parts:

  1. The dynamics of behavior and cognition
    In this part of the course you are introduced some examples of human behavior that challenge the assumptions of linear statistics: reading time, voice dynamics in clinical populations, etc. You are then shown step-by-step how to characterize and quantify patterns and temporal dynamics in these behaviors using non-linear methods, such as recurrence quantification analysis.
  2. The dynamics of interpersonal coordination
    In this second part of the course we focus on interpersonal coordination: how do people manage to coordinate action, emotion and cognition? We consider several real-world cases: heart beat synchronization during firewalking rituals, voice adaptation during conversations, joint problem solving in creative tasks – such as building Lego models together. You are then shown ways to analyze how two or more behaviors are coordinated and how to characterize their coupling – or lack-thereof.

This course provides a theoretical and practical introduction to non-linear techniques for social and cognitive sciences. It presents concrete case studies from actual research projects on human behavior and cognition. It encourages you to put all this to practice via practical exercises and quizzes. By the end of this course you will be fully equipped to go out and do your own research projects applying non-linear methods on human behavior and coordination.

Learning objectives

  • Given a timeseries (e.g. a speech recording, or a sequence of reaction times), characterize its patterns: does it contain repetitions? How stable? How complex?
  • Given a timeseries (e.g. a speech recording, or a sequence of reaction times), characterize how it changes over time.
  • Given two timeseries (e.g. the movements of two dancers) characterize their coupling: how do they coordinate? Do they become more similar over time? Can you identify who is leading and who is following?

MOOC relevance

Social and cognitive research is increasingly investigating phenomena that are temporally unfolding and non-linear. However, most educational institutions only offer courses in linear statistics for social scientists. Hence, there is a need for an easy to understand introduction to non-linear analytical tools in a way that is specifically aimed at social and cognitive sciences. The combination of actual cases and concrete tools to analyze them will give the course a wide appeal.
Additionally, methods oriented courses on MOOC platforms such as Coursera have generally proved very attractive for students.

Please spread the word about this interesting course!

Twitter Follow-up: Can MVPA Invalidate Simulation Theory?

Thanks to the wonders of social media, while I was out grocery shopping I received several interesting and useful responses to my previous post on the relationship between multivariate pattern analysis and simulation theory. Rather than try and fit my responses into 140 characters, I figured i’d take a bit more space here to hash them out. I think the idea is really enhanced by these responses, which point to several findings and features of which I was not aware. The short answer seems to be, no MVPA does not invalidate simulation theory (ST) and may even provide evidence for it in the realm of motor intentions, but that we might be able to point towards a better standard of evidence for more exploratory applications of ST (e.g. empathy-for-pain). An important point to come out of these responses as one might expect, is that the interpretation of these methodologies is not always straightforward.

I’ll start with Antonia Hamilton’s question, as it points to a bit of literature that speaks directly to the issue:

antonio_reply

Antonia is referring to this paper by Oosterhof and colleagues, where they directly compare passive viewing and active performance of the same paradigm using decoding techniques. I don’t read nearly as much social cognition literature as I used to, and wasn’t previously aware of this paper. It’s really a fascinating project and I suggest anyone interested in this issue read it at once (it’s open access, yay!). In the introduction the authors point out that spatial overlap alone cannot demonstrate equivalent mechanisms for viewing and performing the same action:

Numerous functional neuroimaging studies have identified brain regions that are active during both the observation and the execution of actions (e.g., Etzel et al. 2008; Iacoboni et al. 1999). Although these studies show spatial overlap of frontal and parietal activations elicited by action observation and execution, they do not demonstrate representational overlap between visual and motor action representations. That is, spatially overlapping activations could reflect different neural populations in the same broad brain regions (Gazzola and Keysers 2009; Morrison and Downing 2007; Peelen and Downing 2007b). Spatial overlap of activations per se cannot establish whether the patterns of neural response are similar for a given action (whether it is seen or performed) but different for different actions, an essential property of the “mirror system” hypothesis.”

They then go on to explain that while MVPA could conceivably demonstrate a simulation-like mechanism (i.e. a common neural representation for viewing/doing), several previous papers attempting to show just that failed to do so. The authors suggest that this may be due to a variety of methodological limitations, which they set out to correct for in their JNPhys publication. Oosterhof et al show that clusters of voxels located primarily in the intraparietal and superior temporal sulci encode cross-modal information, that is code similar information both when viewing and doing:

Click to go to PDF.

From Oosterhof et al, showing combined classification accuray for (train see, test do; train do, test see).

Essentially Oosterhof et al trained their classifier on one modality (see or do) , tested the classifier on the opposite modality in another session, and then repeated this procedure for all possible combinations of session and modality (while appropriately correcting for multiple comparisons). The map above represents the combined classification accuracy from both train-test combinations; interestingly in the supplementary info they show that the maps do slightly differ depend on what was trained:

Click to go to SI.

From supplementary info, A shows classifier trained on see, tested on do, B shows the opposite.

Oosterhof and colleagues also investigate the specificity of information for particular gestures in a second experiment, but for our purposes lets focus on just the first. My first thought is that this does actually provide some evidence for a simulation theory of understanding motor intentions. Clearly there is enough information in each modality to accurately decode the opposite modality: there are populations of neurons encoding similar information both for action execution and perception. Realistically I think this has to be the minimal burden of proof needed to consider an imaging finding to be evidence for simulation theory. So the results of Oosterhof et al do provide supporting evidence for simulation theory in the domain of motor intentions.

Nonetheless, the results also strengthen the argument that more exploratory extentions of ST (like empathy-for-pain) must be held to a similar burden of proof before generalization in these domains is supported. Simply showing spatial overlap is not evidence of simulation, as Oosterhof themselves argue. I think it is interesting to note the slight spatial divergence between the two train-test maps (see on do, do on see). While we can obviously identify voxels encoding cross-modality information, it is interesting that those voxels do not subsume the entirety of whatever neural computation relates these two modalities; each has something unique to predict in the other. I don’t think that observation invalidates simulation theory, but it might suggest an interesting mechanism not specified in the ‘vanilla’ flavor of ST. To be extra boring, it would be really nice to see an independent replication of this finding, since as Oosterhof themselves point out, the evidence for cross-modal information is inconsistent across studies. Even though the classifier performs well above chance in this study,  it is also worth noting that the majority of surviving voxels in their study show somewhere around 40-50% classification accuracy, not exactly gangbusters. It would be interesting to see if they could identify voxels within these regions that selectively encode only viewing or performing; this might be evidence for a hybrid-theory account of motor intentions.

leoreply

Leonhard’s question is an interesting one that I don’t have a ready response for. As I understand it, the idea is that demonstrating no difference of patterns between a self and other-related condition (e.g. performing an action vs watching someone else do it) might actually be an argument for simulation, since this could be caused by that region using isomorphic computations for both conditions. This an interesting point – i’m not sure what the status of null findings is in the decoding literature, but this merits further thought.

The next two came from James Kilner and Tal Yarkoni. I’ve put them together as I think they fall under a more methodological class of questions/comments and I don’t feel quite experienced enough to answer them- but i’d love to hear from someone with more experience in multivariate/multivoxel techniques:

kilner_reply

talreply

James Kilner asks about the performance of MVPA in the case that the pattern might be spatially overlapping but not identical for two conditions. This is an interesting question and i’m not sure I know the correct answer; my intuition is that you could accurately discriminate both conditions using the same voxels and that this would be strong evidence against a simple simulation theory account (spatial overlap but representational heterogeneity).

Here is more precise answer to James’ question from Sam Schwarzkopf, posted in the comments of the original post:

2. The multivariate aspect obviously adds sensitivity by looking at pattern information, or generally any information of more than one variable (e.g. voxels in a region). As such it is more sensitive to the information content in a region than just looking at the average response from that region. Such an approach can reveal that region A contains some diagnostic information about an experimental variable while region B does not, even though they both show the same mean activation. This is certainly useful knowledge that can help us advance our understanding of the brain – but in the end it is still only one small piece in the puzzle. And as both Tal and James pointed out (in their own ways) and as you discussed as well, you can’t really tell what the diagnostic information actually represents.
Conversely, you can’t be sure that just because MVPA does not pick up diagnostic information from a region that it therefore doesn’t contain any information about the variable of interest. MVPA can only work as long as there is a pattern of information within the features you used.

This last point is most relevant to James’ comment. Say you are using voxels as features to decode some experimental variable. If all the neurons with different tuning characteristics in an area are completely intermingled (like orientation-preference in mouse visual cortex for instance) you should not really see any decoding – even if the neurons in that area are demonstrably selective to the experimental variable.

In general it is clear that the interpretation of decoded patterns is not straightforward- it isn’t clear precisely what information they reflect, and it seems like if a region contained a totally heterogeneous population of neurons you wouldn’t pick up any decoding at all. With respect to ST,  I don’t know if this completely invalidates our ability to test predictions- I don’t think one would expect such radical heterogeneity in a region like STS, but rather a few sub-populations responding selectively to self and other, which MVPA might be able to reveal. It’s an important point to consider though.

Tal’s point is an important one regarding the different sources of information that GLM and MVPA techniques pick up. The paper he refers to by Jimura and Poldrack set out to investigate exactly this by comparing the spatial conjunction and divergent sensitivity of each method. Importantly they subtracted the mean of each beta-coefficient from the multivariate analysis to insure that the analysis contained only information not in the GLM:

pold_mvpa

As you can see in the above, Jimura and Poldrack show that MVPA picks up a large number of voxels not found in the GLM analysis. Their interpretation is that the GLM is designed to pick up regions responding globally or in most cases to stimulation, whereas MVPA likely picks up globally distributed responses that show variance in their response. This is a bit like the difference between functional integration and localization; both are complementary to the understanding of some cognitive function. I take Tal’s point to be that the MVPA and GLM are sensitive to different sources of information and that this blurs the ability of the technique to evaluate simulation theory- you might observe differences between the two that would resemble evidence against ST (different information in different areas) when in reality you would be modelling altogether different aspects of the cognition. edit: after more discussion with Tal on Twitter, it’s clear that he meant to point out the ambiguity inherent in interpreting the predictive power of MVPA; by nature these analyses will pick up a lot of confounding a causal noise- arousal, reaction time, respiration, etc, which would be excluded in a GLM analysis. So these are not necessarily or even likely to be “direct read-outs” of representations, particularly to the extent that such confounds correlate with the task. See this helpful post by neuroskeptic for an overview of one recent paper examining this issue. See here for a study investigating the complex neurovascular origins of MVPA for fMRI. 

Thanks sincerely for these responses, as it’s been really interesting and instructive for me to go through these papers and think about their implications. I’m still new to these techniques and it is exciting to gain a deeper appreciation of the subtleties involved in their interpretation. On that note, I must direct you to check out Sam Schwarzkopf’s excellent reply to my original post. Sam points out some common misunderstandings (of which I am perhaps guilty of several) regarding the interpretation of MVPA/decoding versus GLM techniques, arguing essentially that they pick up much of the same information and can both be considered ‘decoding’ in some sense, further muddying their ability to resolves debates like that surrounding simulation theory.

Will multivariate decoding spell the end of simulation theory?

Decoding techniques such as multivariate pattern analysis (MVPA) are hot stuff in cognitive neuroscience, largely because they offer a tentative promise of actually reading out the underlying computations in a region rather than merely describing data features (e.g. mean activation profiles). While I am quite new to MVPA and similar machine learning techniques (so please excuse any errors in what follows), the basic process has been explained to me as a reversal of the X and Y variables in a typical general linear model. Instead of specifying a design matrix of explanatory (X) variables and testing how well those predict a single independent (Y) variable (e.g. the BOLD timeseries in each voxel), you try to estimate an explanatory variable (essentially decoding the ‘design matrix’ that produced the observed data) from many Y variables, for example one Y variable per voxel (hence the multivariate part). The decoded explanatory variable then describes (BOLD) responses in way that can vary in space, rather than reflecting an overall data feature across a set of voxels such as mean or slope. Typically decoding analyses proceed in two steps, one in which you train the classifier on some set of voxels and another where you see how well that trained model can classify patterns of activity in another scan or task. It is precisely this ability to detect patterns in subtle spatial variations that makes MVPA an attractive technique- the GLM simply doesn’t account for such variation.

The implicit assumption here is that by modeling subtle spatial variations across a set of voxels, you can actually pick up the neural correlates of the underlying computation or representation (Weil and Rees, 2010, Poldrack, 2011). To illustrate the difference between an MVPA and GLM analysis, imagine a classical fMRI experiment where we have some set of voxels defining a region with a significant mean response to your experimental manipulation. All the GLM can tell us is that in each voxel the mean response is significantly different from zero. Each voxel within the significant region is likely to vary slightly in its actual response- you might imagine all sorts of subtle intensity variations within a significant region- but the GLM essentially ignores this variation. The exciting assumption driving interest in decoding is that this variability might actually reflect the activity of sub-populations of neurons and by extension, actual neural representations. MVPA and similar techniques are designed to pick out when these reflect a coherent pattern; once identified this pattern can be used to “predict” when the subject was seeing one or another particular stimulus. While it isn’t entirely straightforward to interpret the patterns MVPA picks out as actual ‘neural representations’, there is some evidence that the decoded models reflect a finer granularity of neural sub-populations than represented in overall mean activation profiles (Todd, 2013; Thompson 2011).

Professor Xavier applies his innate talent for MVPA.

Professor Xavier applies his innate talent for MVPA.

As you might imagine this is terribly exciting, as it presents the possibility to actually ‘read-out’ the online function of some brain area rather than merely describing its overall activity. Since the inception of brain scanning this has been exactly the (largely failed) promise of imaging- reverse inference from neural data to actual cognitive/perceptual contents. It is understandable then that decoding papers are the ones most likely to appear in high impact journals- just recently we’ve seen MVPA applied to dream states, reconstruction of visual experience, and pain experience all in top journals (Kay et al., 2008, Horikawa et al., 2013, Wager et al., 2013). I’d like to focus on that last one for the remainer of this post, as I think we might draw some wide-reaching conclusions for theoretical neuroscience as a whole from Wager et al’s findings.

Francesca and I were discussing the paper this morning- she’s working on a commentary for a theoretical paper concerning the role of the “pain matrix” in empathy-for-pain research. For those of you not familiar with this area, the idea is a basic simulation-theory argument-from-isomorphism. Simulation theory (ST) is just the (in)famous idea that we use our own motor system (e.g. mirror neurons) to understand the gestures of others. In a now infamous experiment Rizzolatti et al showed that motor neurons in the macaque monkey responded equally to their own gestures or the gestures of an observed other (Rizzolatti and Craighero, 2004). They argued that this structural isomorphism might represent a general neural mechanism such that social-cognitive functions can be accomplished by simply applying our own neural apparatus to work out what was going on for the external entity. With respect to phenomena such empathy for pain and ‘social pain’ (e.g. viewing a picture of someone you broke up with recently), this idea has been extended to suggest that, since a region of networks known as “the pain matrix” activates similarly when we are in pain or experience ‘social pain’, that we “really feel” pain during these states (Kross et al., 2011) [1].

In her upcoming commentary, Francesca points out an interesting finding in the paper by Wager and colleagues that I had overlooked. Wager et al apply a decoding technique in subjects undergoing painful and non-painful stimulation. Quite impressively they are then able to show that the decoded model predicts pain intensity in different scanners and various experimental manipulations. However they note that the model does not accurately predict subject’s ‘social pain’ intensity, even though the subjects did activate a similar network of regions in both the physical and social pain tasks (see image below). One conclusion from these findings it that it is surely premature to conclude that because a group of subjects may activate the same regions during two related tasks, those isomorphic activations actually represent identical neural computations [2]. In other words, arguments from structural isomorpism like ST don’t provide any actual evidence for the mechanisms they presuppose.

Figure from Wager et al demonstrating specificity of classifier for pain vs warmth and pain vs rejection. Note poor receiver operating curve (ROC) for 'social pain' (rejecter vs friend), although that contrast picks out similar regions of the 'pain matrix'.

Figure from Wager et al demonstrating specificity of classifier for pain vs warmth and pain vs rejection. Note poor receiver operating curve (ROC) for ‘social pain’ (rejecter vs friend), although that contrast picks out similar regions of the ‘pain matrix’.

To me this is exactly the right conclusion to take from Wager et al and similar decoding papers. To the extent that the assumption that MVPA identifies patterns corresponding to actual neural representations holds, we are rapidly coming to realize that a mere mean activation profile tells us relatively little about the underlying neural computations [3]. It certainly does not tell us enough to conclude much of anything on the basis that a group of subjects activate “the same brain region” for two different tasks. It is possible and even likely that just because I activate my motor cortex when viewing you move, I’m doing something quite different with those neurons than when I actually move about. And perhaps this was always the problem with simulation theory- it tries to make the leap from description (“similar brain regions activate for X and Y”) to mechanism, without actually describing a mechanism at all. I guess you could argue that this is really just a much fancier argument against reverse inference and that we don’t need MVPA to do away with simulation theory. I’m not so sure however- ST remains a strong force in a variety of domains. If decoding can actually do away with ST and arguments from isomorphism or better still, provide a reasonable mechanism for simulation, it’ll be a great day in neuroscience. One thing is clear- model based approaches will continue to improve cognitive neuroscience as we go beyond describing what brain regions activate during a task to actually explaining how those regions work together to produce behavior.

I’ve curated some enlightening responses to this post in a follow-up – worth checking for important clarifications and extensions! See also the comments on this post for a detailed explanation of MVPA techniques. 

References

Horikawa T, Tamaki M, Miyawaki Y, Kamitani Y (2013) Neural Decoding of Visual Imagery During Sleep. Science.

Kay KN, Naselaris T, Prenger RJ, Gallant JL (2008) Identifying natural images from human brain activity. Nature 452:352-355.

Kross E, Berman MG, Mischel W, Smith EE, Wager TD (2011) Social rejection shares somatosensory representations with physical pain. Proceedings of the National Academy of Sciences 108:6270-6275.

Poldrack RA (2011) Inferring mental states from neuroimaging data: from reverse inference to large-scale decoding. Neuron 72:692-697.

Rizzolatti G, Craighero L (2004) The mirror-neuron system. Annu Rev Neurosci 27:169-192.

Thompson R, Correia M, Cusack R (2011) Vascular contributions to pattern analysis: Comparing gradient and spin echo fMRI at 3T. Neuroimage 56:643-650.

Todd MT, Nystrom LE, Cohen JD (2013) Confounds in Multivariate Pattern Analysis: Theory and Rule Representation Case Study. NeuroImage.

Wager TD, Atlas LY, Lindquist MA, Roy M, Woo C-W, Kross E (2013) An fMRI-Based Neurologic Signature of Physical Pain. New England Journal of Medicine 368:1388-1397.

Weil RS, Rees G (2010) Decoding the neural correlates of consciousness. Current opinion in neurology 23:649-655.


[1] Interestingly this paper comes from the same group (Wager et al) showing that pain matrix activations do NOT predict ‘social’ pain. It will be interesting to see how they integrate this difference.

[2] Nevermind the fact that the ’pain matrix’ is not specific for pain.

[3] With all appropriate caveats regarding the ability of decoding techniques to resolve actual representations rather than confounding individual differences (Todd et al., 2013) or complex neurovascular couplings (Thompson et al., 2011).

Mental Training and Neuroplasticity – PhD Complete!

I was asked to write a brief summary of my PhD research for our annual CFIN report. I haven’t blogged in a while and it turned out to be a decent little blurb, so I figured I might as well share it here. Enjoy!

In the past decade, reports concerning the natural plasticity of the human brain have taken a spotlight in the media and popular imagination. In the pursuit of neural plasticity nearly every imaginable specialization, from taxi drivers to Buddhist monks, has had their day in the scanner. These studies reveal marked functional and structural neural differences between various populations of interest, and in doing so drive a wave of interest in harnessing the brain’s plasticity for rehabilitation, education, and even increasing intelligence (Green and Bavelier, 2008). Under this new “mental training” research paradigm investigators are now examining what happens to brain and behavior when novices are randomized to a training condition, using longitudinal brain imaging.

Image1_training

These studies highlight a few promising domains for harnessing neural plasticity, particularly in the realm of visual attention, cognitive control, and emotional training. By randomizing novices to a brief ‘dose’ of action video game or meditation training, researchers can go beyond mere cross-section and make inferences regarding the causality of training on observed neural outcomes. Initial results are promising, suggesting that domains of great clinical relevance such as emotional and attentional processing are amenable to training (Lutz et al., 2008a; Lutz et al., 2008b; Bavelier et al., 2010). However, these findings are currently obscured by a host of methodological limitations.

These span from behavioral confounds (e.g. motivation and demand characteristic) to inadequate longitudinal processing of brain images, which present particular challenges not found in within-subject or cross-sectional design (Davidson, 2010; Jensen et al., 2011). The former can be addressed directly by careful construction of “active control” groups. Here both comparison and control groups receive putatively effective treatments, carefully designed to isolate the hypothesized “active-ingredients” involved in behavioral and neuroplasticity outcomes. In this way researchers can simultaneously make inferences in terms of mechanistic specificity while excluding non-specific confounds such as social support, demand, and participant motivation.

image2_meditationbrainWe set out to investigate one particularly popular intervention, mindfulness meditation, while controlling for these factors. Mindfulness meditation has enjoyed a great deal of research interest in recent years. This popularity is largely due to promising findings indicating good efficacy of meditation training (MT) for emotion processing and cognitive control (Sedlmeier et al., 2012). Clinical studies indicate that MT may be particularly effective for disorders that are typically non-responsive to cognitive-behavioral therapy, such as severe depression and anxiety (Grossman et al., 2004; Hofmann et al., 2010). Understanding the neural mechanism underlying such benefits remains difficult however, as most existing investigations are cross-sectional in nature or depend upon inadequate “wait-list” passive control groups.

We addressed these difficulties in an investigation of functional and structural neural plasticity before and after a 6-week active-controlled mindfulness intervention. To control demand, social support, teacher enthusiasm, and participant motivation we constructed a “shared reading and listening” active control group for comparison to MT. By eliciting daily “experience samples” regarding participants’ motivation to practice and minutes practiced, we ensured that groups did not differ on common motivational confounds.

We found that while both groups showed equivalent improvement on behavioral response-inhibition and meta-cognitive measures, only the MT group significantly reduced affective-Stroop conflict reaction times (Allen et al., 2012). Further we found that MT participants show significantly greater increases in recruitment of dorsolateral prefrontal cortex than did controls, a region implicated in cognitive control and working memory. Interestingly we did not find group differences in emotion-related reaction times or BOLD activity; instead we found that fronto-insula and medial-prefrontal BOLD responses in the MT group were significantly more correlated with practice than in controls. These results indicate that while brief MT is effective for training attention-related neural mechanisms, only participants with the greatest amount of practice showed altered neural responses to negative affective stimuli. This result is important because it underlines the differential response of various target skills to training and suggests specific applications of MT depending on time and motivation constraints.

MT related increase in DLPFC activity during affective stroop task.

MT related increase in DLPFC activity during affective stroop task.

In a second study, we utilized a longitudinally optimized pipeline to assess structural neuroplasticity in the same cohort as described above (Ashburner and Ridgway, 2012). A crucial issue in longitudinal voxel-based morphometry and similar methods is the prevalence of “asymmetrical preprocessing”, for example where normalization parameters are calculated from baseline images and applied to follow-up images, resulting in inflated risk of false-positive results. We thus applied a totally symmetrical deformation-based morphometric pipeline to assess training related expansions and contractions of gray matter volume. While we found significant increases within the MT group, these differences did not survive group-by-time comparison and thus may represent false positives; it is likely that such differences would not be ruled out by an asymmetric pipeline or non-active controlled designed. These results suggest that brief MT may act only on functional neuroplasticity and that greater training is required for more lasting anatomical alterations.

These projects are a promising advance in our understanding of neural plasticity and mental training, and highlight the need for careful methodology and control when investigating such phenomena. The investigation of neuroplasticity mechanisms may one day revolutionize our understanding of human learning and neurodevelopment, and we look forward to seeing a new wave of carefully controlled investigations in this area.

You can read more about the study in this blog post, where I explain it in detail. 

A happy day, my PhD defense!

A happy day, my PhD defense!

References

Allen M, Dietz M, Blair KS, van Beek M, Rees G, Vestergaard-Poulsen P, Lutz A, Roepstorff A (2012) Cognitive-Affective Neural Plasticity following Active-Controlled Mindfulness Intervention. The Journal of Neuroscience 32:15601-15610.

Ashburner J, Ridgway GR (2012) Symmetric diffeomorphic modeling of longitudinal structural MRI. Frontiers in neuroscience 6.

Bavelier D, Levi DM, Li RW, Dan Y, Hensch TK (2010) Removing brakes on adult brain plasticity: from molecular to behavioral interventions. The Journal of Neuroscience 30:14964-14971.

Davidson RJ (2010) Empirical explorations of mindfulness: conceptual and methodological conundrums. Emotion 10:8-11.

Green C, Bavelier D (2008) Exercising your brain: a review of human brain plasticity and training-induced learning. Psychology and Aging; Psychology and Aging 23:692.

Grossman P, Niemann L, Schmidt S, Walach H (2004) Mindfulness-based stress reduction and health benefits: A meta-analysis. Journal of Psychosomatic Research 57:35-43.

Hofmann SG, Sawyer AT, Witt AA, Oh D (2010) The effect of mindfulness-based therapy on anxiety and depression: A meta-analytic review. Journal of consulting and clinical psychology 78:169.

Jensen CG, Vangkilde S, Frokjaer V, Hasselbalch SG (2011) Mindfulness training affects attention—or is it attentional effort?

Lutz A, Brefczynski-Lewis J, Johnstone T, Davidson RJ (2008a) Regulation of the neural circuitry of emotion by compassion meditation: effects of meditative expertise. PLoS One 3:e1897.

Lutz A, Slagter HA, Dunne JD, Davidson RJ (2008b) Attention regulation and monitoring in meditation. Trends Cogn Sci 12:163-169.

Sedlmeier P, Eberth J, Schwarz M, Zimmermann D, Haarig F, Jaeger S, Kunze S (2012) The psychological effects of meditation: A meta-analysis.

Active-controlled, brief body-scan meditation improves somatic signal discrimination.

Here in the science blog-o-sphere we often like to run to the presses whenever a laughably bad study comes along, pointing out all the incredible feats of ignorance and sloth. However, this can lead to science-sucks cynicism syndrome (a common ailment amongst graduate students), where one begins to feel a bit like all the literature is rubbish and it just isn’t worth your time to try and do something truly proper and interesting. If you are lucky, it is at this moment that a truly excellent paper will come along at the just right time to pick up your spirits and re-invigorate your work. Today I found myself at one such low-point, struggling to figure out why my data suck, when just such a beauty of a paper appeared in my RSS reader.

data_sensing (1)The paper, “Brief body-scan meditation practice improves somatosensory perceptual decision making”, appeared in this month’s issue of Consciousness and Cognition. Laura Mirams et al set out to answer a very simple question regarding the impact of meditation training (MT) on a “somatic signal detection task” (SSDT). The study is well designed; after randomization, both groups received audio CDs with 15 minutes of daily body-scan meditation or excerpts from The Lord of The Rings. For the SSD task, participants simply report when they felt a vibration stimulus on the finger, where the baseline vibration intensity is first individually calibrated to a 50% detection rate. The authors then apply a signal-detection analysis framework to discern the sensitivity or d’ and decision criteria c.

Mirams et al found that, even when controlling for a host of baseline factors including trait mindfulness and baseline somatic attention, MT led to a greater increase in d’ driven by significantly reduced false-alarms. Although many theorists and practitioners of MT suggest a key role for interoceptive & somatic attention in related alterations of health, brain, and behavior, there exists almost no data addressing this prediction, making these findings extremely interesting. The idea that MT should impact interoception and somatosensation is very sensible- in most (novice) meditation practices it is common to focus attention to bodily sensations of, for example, the breath entering the nostril. Further, MT involves a particular kind of open, non-judgemental awareness of bodily sensations, and in general is often described to novice students as strengthening the relationship between the mind and sensations of the body. However, most existing studies on MT investigate traditional exteroceptive, top-down elements of attention such as conflict resolution and the ability to maintain attention fixation for long periods of time.

While MT certainly does involve these features, it is arguable that the interoceptive elements are more specific to the precise mechanisms of interest (they are what you actually train), whereas the attentional benefits may be more of a kind of side effect, reflecting an early emphasis in MT on establishing attention. Thus in a traditional meditation class, you might first learn some techniques to fixate your attention, and then later learn to deploy your attention to specific bodily targets (i.e. the breath) in a particular way (non-judgmentally). The goal is not necessarily to develop a super-human ability to filter distractions, but rather to change the way in which interoceptive responses to the world (i.e. emotional reactions) are perceived and responded to. This hypothesis is well reflected in the elegant study by Mirams et al; they postulate specifically that MT will lead to greater sensitivity (d’), driven by reduced false alarms rather than an increased hit-rate, reflecting a greater ability to discriminate the nature of an interoceptive signal from noise (note: see comments for clarification on this point by Steve Fleming – there is some ambiguity in interpreting the informational role of HR and FA in d’). This hypothesis not only reflects the theoretically specific contribution of MT (beyond attention training, which might be better trained by video games for example), but also postulates a mechanistically specific hypothesis to test this idea, namely that MT leads to a shift specifically in the quality of interoceptive signal processing, rather than raw attentional control.

At this point, you might ask if everyone is so sure that MT involves training interoception, why is there so little data on the topic? The authors do a great job reviewing findings (even including currently in-press papers) on interoception and MT. Currently there is one major null finding using the canonical heartbeat detection task, where advanced practitioners self-reported improved heart beat detection but in reality performed at chance. Those authors speculated that the heartbeat task might not accurately reflect the modality of interoception engaged in by practitioners. In addition a recent study investigated somatic discrimination thresholds in a cross-section of advanced practitioners and found that the ability to make meta-cognitive assessments of ones’ threshold sensitivity correlated with years of practice. A third recent study showed greater tactile sensation acuity in practitioners of Tai Chi.  One longitudinal study [PDF], a wait-list controlled fMRI investigation by Farb et al, found that a mindfulness-based stress reduction course altered BOLD responses during an attention-to-breath paradigm. Collectively these studies do suggest a role of MT in training interoception. However, as I have complained of endlessly, cross-sections cannot tell us anything about the underlying causality of the observed effects, and longitudinal studies must be active-controlled (not waitlisted) to discern mechanisms of action. Thus active-controlled longitudinal designs are desperately needed, both to determine the causality of a treatment on some observed effect, and to rule out confounds associated with motivation, demand-characteristic, and expectation. Without such a design, it is very difficult to conclude anything about the mechanisms of interest in an MT intervention.

In this regard, Mirams went above and beyond the call of duty as defined by the average paper. The choice of delivering the intervention via CD is excellent, as we can rule out instructor enthusiasm/ability confounds. Further the intervention chosen is extremely simple and well described; it is just a basic body-scan meditation without additional fluff or fanfare, lending to mechanistic specificity. Both groups were even instructed to close their eyes and sit when listening, balancing these often overlooked structural factors. In this sense, Mirams et al have controlled for instruction, motivation, intervention context, baseline trait mindfulness, and even isolated the variable of interest- only the MT group worked with interoception, though both exerted a prolonged period of sustained attention. Armed with these controls we can actually say that MT led to an alteration in interoceptive d’, through a mechanism dependent upon on the specific kind of interoceptive awareness trained in the intervention.

It is here that I have one minor nit-pick of the paper. Although the use of Lord of the Rings audiotapes is with precedent, and likely a great control for attention and motivation, you could be slightly worried that reading about Elves and Orcs is not an ideal control for listening to hours of tapes instructing you to focus on your bodily sensations, if the measure of interest involves fixating on the body. A pure active control might have been a book describing anatomy or body parts; then we could exhaustively conclude that not only is it interoception driving the findings, but the particular form of interoceptive attention deployed by meditation training. As it is, a conservative person might speculate that the observed differences reflect demand characteristics- MT participants deploy more attention to the body due to a kind of priming mechanism in the teaching. However this is an extreme nitpick and does not detract from the fact that Mirams and co-authors have made an extremely useful contribution to the literature. In the future it would be interesting to repeat the paradigm with a more body-oriented control, and perhaps also in advanced practitioners before and after an intensive retreat to see if the effect holds at later stages of training. Of course, given my interest in applying signal-detection theory to interoceptive meta-cognition, I also cannot help but wonder what the authors might have found if they’d applied a Fleming-style meta-d’ analysis to this study.

All in all, a clear study with tight methods, addressing a desperately under-developed research question, in an elegant fashion. The perfect motivation to return to my own mangled data ☺

Correcting your naughty insula: modelling respiration, pulse, and motion artifacts in fMRI

important update: Thanks to commenter “DS”, I discovered that my respiration-related data was strongly contaminated due to mechanical error. The belt we used is very susceptible to becoming uncalibrated, if the subject moves or breathes very deeply for example. When looking at the raw timecourse of respiration I could see that many subjects, included the one displayed here, show a great deal of “clipping” in the timeseries. For the final analysis I will not use the respiration regressors, but rather just the pulse and motion. Thanks DS!

As I’m working my way through my latest fMRI analysis, I thought it might be fun to share a little bit of that here. Right now i’m coding up a batch pipeline for data from my Varela-award project, in which we compared “adept” meditation practitioners with motivation, IQ, age, and gender-matched controls on a response-inhibition and error monitoring task. One thing that came up in the project proposal meeting was a worry that, since meditation practitioners spend so much time working with the breath, they might respirate differently either at rest or during the task. As I’ve written about before, respiration and other related physiological variables such as cardiac-pulsation induced motion can seriously impact your fMRI results (when your heart beats, the veins in your brain pulsate, creating slight but consistent and troublesome MR artifacts). As you might expect, these artifacts tend to be worse around the main draining veins of the brain, several of which cluster around the frontoinsular and medial-prefrontal/anterior cingulate cortices. As these regions are important for response-inhibition and are frequently reported in the meditation literature (without physiological controls), we wanted to try to control for these variables in our study.

disclaimer: i’m still learning about noise modelling, so apologies if I mess up the theory/explanation of the techniques used! I’ve left things a bit vague for that reason. See bottom of article for references for further reading. To encourage myself to post more of these “open-lab notes” posts, I’ve kept the style here very informal, so apologies for typos or snafus. :D

To measure these signals, we used the respiration belt and pulse monitor that come standard with most modern MRI machines. The belt is just a little elastic hose that you strap around the chest wall of the subject, where it can record expansions and contractions of the chest to give a time series corresponding to respiration, and the pulse monitor a standard finger clip. Although I am not an expert on physiological noise modelling, I will do my best to explain the basic effects you want to model out of your data. These “non-white” noise signals include pulsation and respiration-induced motion (when you breath, you tend to nod your head just slightly along the z-axis), typical motion artifacts, and variability of pulsation and respiration. To do this I fed my physiological parameters into an in-house function written by Torben Lund, which incorporates a RETROICOR transformation of the pulsation and respiration timeseries. We don’t just use the raw timeseries due to signal aliasing- the phsyio data needs to be shifted to make each physiological event correspond to a TR. The function also calculates the respiratory volume time delay (RVT), a measure developed by Rasmus Birn, to model the variability in physiological parameters1. Variability in respiration and pulse volume (if one group of subjects tend to inhale sharply for some conditions but not others, for example) is more likely to drive BOLD artifacts than absolute respiratory volume or frequency (if one group of subjects tend to inhale sharply for some conditions but not others, for example). Finally, as is standard, I included the realignment parameters to model subject motion-related artifacts. Here is a shot of my monster design matrix for one subject:

DM_NVR

You can see that the first 7 columns model my conditions (correct stops, unaware errors, aware errors, false alarms, and some self-report ratings), the next 20 model the RETROICOR transformed pulse and respiration timeseries, 41 columns for RVT, 6 for realignment pars, and finally my session offsets and constant. It’s a big DM, but since we have over 1000 degrees of freedom, i’m not too worried about all the extra regressors in terms of loss of power. What would be worrisome is if for example stop activity correlated strongly with any of the nuisance variables –  we can see from the orthogonality plot that in this subject at least, that is not the case. Now lets see if we actually have anything interesting left over after we remove all that noise:

stop SPM

We can see that the Stop-related activity seems pretty reasonable, clustering around the motor and premotor cortex, bilateral insula, and DLPFC, all canonical motor inhibition regions (FWE-cluster corrected p = 0.05). This is a good sign! Now what about all those physiological regressors? Are they doing anything of value, or just sucking up our power? Here is the f-contrast over the pulse regressors:

pulse

Here we can see that the peak signal is wrapped right around the pons/upper brainstem. This makes a lot of sense- the area is full of the primary vasculature that ferries blood into and out of the brain. If I was particularly interested in getting signal from the brainstem in this project, I could use a respiration x pulse interaction regressor to better model this6. Penny et al find similar results to our cardiac F-test when comparing AR(1) with higher order AR models [7]. But since we’re really only interested in higher cortical areas, the pulse regressor should be sufficient. We can also see quite a bit of variance explained around the bilateral insula and rostral anterior cingulate. Interestingly, our stop-related activity still contained plenty of significant insula response, so we can feel better that some but not all of the signal from that region is actually functionally relevant. What about respiration?

resp

Here we see a ton of variance explained around the occipital lobe. This makes good sense- we tend to just slightly nod our head back and forth along the z-axis as we breath. What we are seeing is the motion-induced artifact of that rotation, which is most severe along the back of the head and periphery of the brain. We see a similar result for the overall motion regressors, but flipped to the front:

Ignore the above, respiration regressor is not viable due to “clipping”, see note at top of post. Glad I warned everyone that this post was “in progress” :) Respiration should be a bit more global, restricted to ventricles and blood vessels.

motion

Wow, look at all the significant activity! Someone call up Nature and let them know, motion lights up the whole brain! As we would expect, the motion regressor explains a ton of uninteresting variance, particularly around the prefrontal cortex and periphery.

I still have a ways to go on this project- obviously this is just a single subject, and the results could vary wildly. But I do think even at this point we can start to see that it is quite easy and desirable to model these effects in your data (Note: we had some technical failure due to the respiration belt being a POS…) I should note that in SPM, these sources of “non-white” noise are typically modeled using an autoregressive (AR(1)) model, which is enabled in the default settings (we’ve turned it off here). However as there is evidence that this model performs poorly at faster TRs (which are the norm now), and that a noise-modelling approach can greatly improve SnR while removing artifacts, we are likely to get better performance out of a nuisance regression technique as demonstrated here [4]. The next step will be to take these regressors to a second level analysis, to examine if the meditation group has significantly more BOLD variance-explained by physiological noise than do controls. Afterwards, I will re-run the analysis without any physio parameters, to compare the results of both.

References:


1. Birn RM, Diamond JB, Smith MA, Bandettini PA.
Separating respiratory-variation-related fluctuations from neuronal-activity-related fluctuations in fMRI.
Neuroimage. 2006 Jul 15;31(4):1536-48. Epub 2006 Apr 24.

2. Brooks J.C.W., Beckmann C.F., Miller K.L. , Wise R.G., Porro C.A., Tracey I., Jenkinson M.
Physiological noise modelling for spinal functional magnetic resonance imaging studies
NeuroImage in press: DOI: doi: 10.1016/j.neuroimage.2007.09.018

3. Glover GH, Li TQ, Ress D.
Image-based method for retrospective correction of physiological motion effects in fMRI: RETROICOR.
Magn Reson Med. 2000 Jul;44(1):162-7.

4. Lund TE, Madsen KH, Sidaros K, Luo WL, Nichols TE.
Non-white noise in fMRI: does modelling have an impact?
Neuroimage. 2006 Jan 1;29(1):54-66.

5. Wise RG, Ide K, Poulin MJ, Tracey I.
Resting fluctuations in arterial carbon dioxide induce significant low frequency variations in BOLD signal.
Neuroimage. 2004 Apr;21(4):1652-64.

2. Brooks J.C.W., Beckmann C.F., Miller K.L. , Wise R.G., Porro C.A., Tracey I., Jenkinson M.
Physiological noise modelling for spinal functional magnetic resonance imaging studies
NeuroImage in press: DOI: doi: 10.1016/j.neuroimage.2007.09.018

7. Penny, W., Kiebel, S., & Friston, K. (2003). Variational Bayesian inference for fMRI time series. NeuroImage, 19(3), 727–741. doi:10.1016/S1053-8119(03)00071-5

Follow

Get every new post delivered to your Inbox.

Join 11,856 other followers

%d bloggers like this: