Check out this recent editor's pick:
All site content, except where otherwise noted, is licensed under a Creative Commons BY-NC-ND License. Additional permissions for appropriate reuse available on request.
All site content, except where otherwise noted, is licensed under a Creative Commons BY-NC-ND License. Additional permissions for appropriate reuse available on request.
Comments on Related Content
Massieh Moayedi, University of Toronto
This comment was coauthored
This comment was coauthored by David Seminowicz, University of Maryland, Baltimore, US.
Can we use fMRI to determine if someone is in pain and how intense the pain experience is? A recent study in the New England Journal of Medicine by Wager and colleagues argues that we can. They used a pattern recognition technique-based approach to the problem, using one dataset to find a “pain signature” that could predict pain intensity, then using the same algorithm to see if it could distinguish brain patterns related to pain perception from brain patterns related to stimuli that induce other percepts, such as warmth, social rejection (or “social pain”), or opioid-induced analgesia. The pattern analysis successfully parsed these brain activities with a consistent classification accuracy well over 90 percent, and forced choice accuracy—i.e., asking the algorithm to determine if the brain activity was related to pain or to something else—of 100 percent, most of the time. This impressive study was technically sound and well designed, and the results are compelling. It’s easy to appreciate the novelty of the ideas and innovative methods emerging from Wager’s group. There are many more positive things we could say about the study, but we would like to draw attention to some of the limitations, as these may affect how these findings can be effectively translated to patient populations—most of which Wager et al. acknowledge in the paper—which we hope will lead to some good discussions. To us, the paper is limited in the following areas:
Novelty: For two reasons. 1) At least two other papers have been published with similar methods (one by Sean Mackey’s group [Ung et al., 2012], another by Marcus Ploner’s group [Schulz et al., 2012]). Wager and colleagues’ study took it to a new level by using four previously published datasets, and demonstrated better sensitivity and specificity. 2) The major regions in the “pain signature” are exactly the regions that a pain neuroimaging researcher would expect to see. In fact, they started off by restricting their brain search to regions commonly activated in over 200 previously reported pain imaging studies. So, the study doesn’t actually advance our understanding of brain mechanisms of pain. Additionally, it is predicted that individual regions, such as the anterior insula, have pretty good predictive value alone—it would be nice to see the prediction accuracy for each region of the pain map.
Specificity: The study does not demonstrate that any of the regions are specific to pain, per se. All four studies involved heat pain evoked by a thermode in healthy subjects. They showed that the “pain signature” doesn’t predict “social pain” (arguably a reverse inference-based misnomer; see Yarkoni et al., 2011, and Iannetti and Mouraux, 2011), but the study does not test the specificity of the “pain signature” to other modalities of experimental pain (e.g., inflammatory, cold, electrical, laser-evoked, muscular, visceral), to clinical pains (e.g., back, neuropathic, headache), or to other pain-related percepts (e.g., anticipated pain, fear of pain, imagined pain, air hunger, pain empathy). Additionally, it has been demonstrated that the regions identified as a “pain signature” are not specific to pain, but rather encode the salience of pain (and any other modality—see Downar et al., 2000, 2001; Iannetti and Mouraux, 2010; Legrain et al., 2011; Mouraux et al., 2011). In the study by Wager et al., the painful stimulus is arguably more salient than any of the other stimuli, and salience was not assessed. Therefore, the pattern analysis may only be detecting what we—the pain imaging research community—have previously demonstrated with standard (univariate) statistical methods: that pain is more salient than warm, social rejection, etc.
Clinical relevance: We believe neuroimaging has incredible potential to influence clinical practice and development of new treatments, and we think this paper has translational potential. But this technology is certainly not ready for the clinic. Wager et al. concede this point, but the choice of journals is questionable: NEJM is considered a top, evidence-based medicine journal, so how does this study inform clinicians with regard to their day-to-day clinical practice? (See the editorials and commentaries in the April 2013 issue of the Journal of Pain for some discussions about the using neuroimaging for pain assessment).
We’ll end on a more positive note. Outside of the pain world, neuroimaging researchers are working on similar problems, e.g., using fMRI to predict when a person finds different scenes in a movie funny, scary, boring, sexy, confusing, or whatever, or to predict what a person is dreaming about. Wager’s study brings pain research to the cutting edge of the field. Wager et al. have made all the tools necessary to replicate their methods freely available. Thus, a great opportunity exists to test this “pain signature” on many other datasets. We say, let’s test it.
Marshall Devor, Hebrew University of Jerusalem
This article makes the
This article makes the radical claim that an objective, neurologic signature of subjective pain experience has been discovered. As this is the Holy Grail of Pain Science, the claim needs to be evaluated closely. In my opinion, nothing even resembling the claim has actually been demonstrated. The basis of this opinion touches on the essence of what we mean by a "brain correlate" of a conscious experience.
Pain is a subjective experience. But barring the supernatural, there must be a specific pattern of brain activity that gives rise to this experience. Such a neural fingerprint would track the pain report of an individual at high fidelity irrespective of intervening parameters, and constitute the driver of conscious experience and (truthful) verbal report. In principle, the pattern ought to be accessible to scientific investigation, although not necessarily with existing equipment. Discovering such neural correlates of pain is a high priority for pain research and for all branches of neuroscience that touch on sensory perception. Progress has been made. For example, particular types of painful stimuli (e.g., heat) yield roughly uniform patterns of activation in a constellation of brain areas called the “pain matrix” (May, 2007). However, this activity could well reflect nothing more than the intensity of the noxious input, something that is very different from pain experience, per se. A key element in proving that brain activity patterns reflect the subjective experience of pain is to dissociate the stimulus strength from the resulting experience. For example, using placebo, hypnotic suggestion, or distraction, pain experience can be substantially changed, even if the input stimulus remains constant. If one is really recording a "pain signature," it ought to co-vary with the pain experience rather than with the input signal. In fact, some cortical areas have been identified at which brain activity correlates with changing pain experience when the applied stimulus was held constant (e.g., Rainville et al. 1997; 1999). However, correlations reported so far tend to be weak and to require multiple tests in multiple individuals. In addition, although the temperature of the heat stimulus applied to the skin in these studies was closely controlled, the authors failed to prove that the afferent input to the brain was invariant. Factors such as sweating, cutaneous blood flow, and spinal gating can modulate the link between skin surface temperature and afferent input to the brain. More work needs to be done.
It is against this background that the announcement (and in the New England Journal of Medicine!) that Tor Wager and coworkers found “An fMRI-based neurological signature of physical pain” is so newsworthy. The Holy Grail is finally in our hands! But reading the paper immediately brings one crashing back to Earth. What the authors actually did was to collect fMRI data from individuals during the application of highly stereotyped heat stimuli that were or were not reported as being painful. They then applied an algorithm, based on machine-learning analysis of the individual subject's data, the output of which correlated highly with that subject’s pain report. The algorithm works by giving specific weights to different components of stimulus-evoked pain matrix activations. The claim of a "pain signature" was based on the fact that the imaging-based correlation was almost as high as the correlation between the pain report and the actual stimulus temperature.
This is perhaps a step forward in the computational analysis of heat-evoked fMRI activations in the brain. But it says absolutely nothing about whether the signal correlates with pain experience. It is more likely that the signal extracted by the algorithm simply monitors the strength of the heat-evoked signal entering the individual’s brain. Does the algorithm provide a “neurological signature of physical pain," or is it no more than a fancy skin thermometer? To know, at a minimum one must dissociate the pain experience from the stimulus strength. Not only was this not done, but the issue was not even raised in the Discussion. The authors correctly pointed out that between their lab finding and real-life application in patients and courts of law, much more needs to be done. But that is not the key weakness of this paper. The authors have, in fact, provided no evidence that the output of their algorithm bears a direct relation to the pain consciously experienced as opposed to the strength of the applied stimulus.
References
Mathieu Roy, UC Boulder
This comment was coauthored
This comment was coauthored by Tor D. Wager, Lauren Y. Atlas, Martin A. Lindquist, Mathieu Roy, Choong-Wan Woo and Ethan Kross.
We thank the Pain Research Forum and the authors of the two commentaries for discussing our paper. We share many of their questions and concerns. Allow us to respond to some of the issues that were raised.
I. Significance & Innovation
Did we find the holy grail for pain neuroimaging research?
No. We did not claim to find a holy grail. In fact, we do not think it is possible to find an objective marker that tracks pain under all circumstances and independent of the processes involved in generating it. One of our hopes is that our research will move the field beyond thinking of pain as a unitary phenomenon (i.e., different types of pain exist).
Is this really novel? Haven't others identified similar regions associated with pain?
Many others have identified the regions we report as part of the "Neurologic pain signature" (NPS), and for good reason: They are all likely related to how the experience of pain is created from acute, noxious input to the brain. However, most studies look for responses to noxious stimuli in single, isolated voxels--and construct brain maps from voxels treated in isolation. What is novel in this context is the idea of developing integrated predictors of pain based on all the available information in the brain maps.
Another challenge to the novelty of this work is based on the notion that previous studies have identified all the regions we report in the NPS map. However, what is a "region?" Many of us think we know what a region is…the anterior mid-cingulate cortex (aMCC) is a "region" that one can localize visually on a brain image. However, the aMCC contains roughly 4,000 mm^3 of brain tissue with on the order of 400 million neurons, with diverse functional properties. Part of the contribution of our paper is to suggest, as many innovators in neuroimaging have before us, that the kind of "region" we can recognize by eye is the wrong level of analysis. The aMCC is shockingly non-specific to physical pain, as many other groups (including us) have pointed out (Yarkoni, Poldrack, Nichols, Van Essen, & Wager, 2011). However, we disagree with the statement that "the regions identified as a ‘pain signature’ are not specific to pain, but rather encode the salience of pain (and any other modality)" because we have identified specific patterns within these regions that do not suffer from these shortcomings in specificity. We have identified a precise pattern within the aMCC that is relatively specific to pain. In conjunction with other patterns in other "regions" that are also somewhat specific to pain, we can differentiate physical pain from a number of other conditions. These patterns have not previously been identified. We believe that a focus on patterns, beyond the simple presence or absence of activations in areas like the aMCC, will improve our understanding of brain function and relate it to behavioral outcomes and psychological states.
Is this completely novel as a concept? How does it differ from previous machine learning approaches in pain research?
No. Other groups have done innovative work in this area (Brodersen et al., 2012; Brown, Chatterjee, Younger, & Mackey, 2011; Cecchi et al., 2012; Liang, Mouraux, Hu, & Iannetti, 2013; Marquand et al., 2010; Schulz, Zherdin, Tiemann, Plant, & Ploner, 2012; Ung et al., 2012; Zhang, Hu, Hung, Mouraux, & Iannetti, 2012). Our research builds on this work by identifying a neural signal that is specific to physical pain, by demonstrating this across studies and scanners, and by demonstrating that this signal does not simply reflect "salience" or general aversion.
Our study also differed from previous attempts (Brodersen et al., 2012; Brown et al., 2011; Liang et al., 2013; Marquand et al., 2010; Schulz et al., 2012; Ung et al., 2012)in both the type of input data and the type of algorithm employed. These choices seem trivial from a distance, but there are important differences in both the experimental design and data analysis approaches that could have substantial impact on the sensitivity and specificity of the resulting brain “signatures.” We expect multiple groups to develop stronger and more specific signatures for different classes of psychological processes as the field matures.
II. Specificity
Does the study demonstrate specificity to pain or not? Is the "neurologic pain signature" really just a salience detector?
We demonstrated specificity to physical pain relative to non-painful warmth, pain anticipation, pain recall, and social pain. That is, we demonstrated that the NPS does not respond to any of these conditions. Furthermore, it does not respond to incremental increases in the intensity of these processes, but it does respond to increases in pain elicited by noxious heat.
Moayedi and Seminowicz argue that "the painful stimulus is arguably more salient than any of the other stimuli, and salience was not assessed." We agree that the painful stimulus may be more salient than the others. It should be noted, however, that there is no single gold-standard measure of "salience," so this is difficult to prove or disprove. We do know that the "social pain" stimuli were rated as equally aversive to the physically painful stimuli.
In addition, our argument rests on more than the comparison between physical pain and pain anticipation, "social pain," etc. We showed that the NPS is not sensitive to the difference between "social pain" and neutral images, and that it was not sensitive to intense vs. mild non-painful warmth. For Moayedi's objection to hold, one would have to posit not only that physical pain is more salient than social pain, but also that the photos of rejectors ("social pain") were not appreciably more salient than photos of friends, in spite of aversion ratings that showed clear differences. One would also have to posit that intense warmth is not more salient than mild warmth, even though behavioral discrimination accuracy for intense vs. mild warmth was comparable to that for intense vs. mild pain. Finally, one would have to posit that cues signaling upcoming pain are not more salient than the fixation cross on-screen between trials. We did not ask participants how "salient" these cues were, but they do reliably induce amygdala activity, aMCC activity, and other brain hallmarks of "salience" and threat. Thus, we think it is likely that the warmth, "social pain," and anticipation conditions we tested were all more salient than their respective controls, and yet they produced no significant effects on the NPS.
We agree that we did not test whether the NPS responds to other modalities of experimental pain (e.g., inflammatory, cold, electrical, laser-evoked, muscular, visceral) or to clinical pains (e.g., back, neuropathic, headache), and that this is an important and potentially fruitful direction for future research. However, we think of this as an issue not of specificity, but of the generalizability of the NPS across pain modalities and types. We also agree that it is important to test the modulatory effects of fear of pain, imagination, cognitive regulation, and other factors. It is also important to test specificity relative to other conditions: air hunger, pain empathy or "observed pain", and anything else that might be confusable for physical pain.
Is the NPS just measuring the intensity of the noxious input?
A related criticism, raised by Devor, is that neurologic pain signature (NPS) activity “could well reflect nothing more than the intensity of the noxious input.” However, this cannot be entirely the case, as we have demonstrated that the NPS signal correlates more strongly with perceived pain than with the intensity of the noxious input.
Another piece of evidence in our study is that the thermal stimuli we use produce a characteristic time course of reported pain due to temporal summation. The stimulus is constant over time, with ramp-up and ramp-down periods. But the pain usually increases continuously, with a characteristic nonlinear shape, and peaks at the time the stimulus turns off. We showed that the time course of the NPS signature during a trial matches the time course of reported pain closely (r > 0.9 for the group average time course), and more closely than the time course of the stimulus itself.
Finally, we showed that the NPS responds strongly to treatment with remifentanil. It might be argued that remifentanil affects spinal gating, and thus the NPS modulation simply reflects reduction of nociceptive input. This is in principle an empirical question, although there is insufficient evidence to know the answer today. What we do know is that a) remifentanil has many effects on supra-spinal processes that correlate with effects on pain; and b) remifentanil likely has a heterogeneous effect on activity across the brain (Atlas et al., 2012; Bingel et al., 2011; Wise et al., 2002; Wise, Williams, & Tracey, 2004). Both these empirical observations argue against a purely spinal mechanism, and thus against the idea that the NPS reduction with remifentanil reflects a reduction in nociceptive input.
Thus, the NPS is not a “fancy skin thermometer,” because a skin thermometer could not account for the likely supraspinal effects that make the NPS track subjective pain more closely than noxious input in both intensity and time, and for opiate effects on the NPS.
Ultimately, we believe there should not be one objective measure of "pain" because "pain" is not one thing. It may not be one experience, but a class of experiences. It is therefore conceivable that under different circumstances, different patterns of brain activity may be associated with identical pain reports. Here, we have developed a pattern that is highly predictive of noxious heat-induced pain. As we and others assess to what extent this pattern generalizes to other types of experimental and clinical pain, and test the effects of various psychological and pharmacological interventions, we may derive a collection of “pain signatures” that together could help identify the processes underlying a given patient’s pain and develop the right treatment for the patient.
References:
Atlas, L. Y., Whittington, R. a, Lindquist, M. a, Wielgosz, J., Sonty, N., & Wager, T. D. (2012). Dissociable influences of opiates and expectations on pain. The Journal of neuroscience : the official journal of the Society for Neuroscience, 32(23), 8053–64. doi:10.1523/JNEUROSCI.0383-12.2012
Bingel, U., Wanigasekera, V., Wiech, K., Ni Mhuircheartaigh, R., Lee, M. C., Ploner, M., & Tracey, I. (2011). The effect of treatment expectation on drug efficacy: imaging the analgesic benefit of the opioid remifentanil. Science translational medicine, 3(70), 70ra14. doi:10.1126/scitranslmed.3001244
Brodersen, K. H., Wiech, K., Lomakina, E. I., Lin, C.-S., Buhmann, J. M., Bingel, U., Ploner, M., et al. (2012). Decoding the perception of pain from fMRI using multivariate pattern analysis. NeuroImage, 63(3), 1162–70. doi:10.1016/j.neuroimage.2012.08.035
Brown, J. E., Chatterjee, N., Younger, J., & Mackey, S. (2011). Towards a physiology-based measure of pain: patterns of human brain activity distinguish painful from non-painful thermal stimulation. PloS one, 6(9), e24124. doi:10.1371/journal.pone.0024124
Cecchi, G. a, Huang, L., Hashmi, J. A., Baliki, M., Centeno, M. V, Rish, I., & Apkarian, a V. (2012). Predictive dynamics of human pain perception. PLoS computational biology, 8(10), e1002719. doi:10.1371/journal.pcbi.1002719
Liang, M., Mouraux, a., Hu, L., & Iannetti, G. D. (2013). Primary sensory cortices contain distinguishable spatial patterns of activity for each sense. Nature Communications, 4(May). doi:10.1038/ncomms2979
Marquand, A., Howard, M., Brammer, M., Chu, C., Coen, S., & Mourão-Miranda, J. (2010). Quantitative prediction of subjective pain intensity from whole-brain fMRI data using Gaussian processes. NeuroImage, 49(3), 2178–89. doi:10.1016/j.neuroimage.2009.10.072
Schulz, E., Zherdin, A., Tiemann, L., Plant, C., & Ploner, M. (2012). Decoding an individual’s sensitivity to pain from the multivariate analysis of EEG data. Cerebral cortex (New York, N.Y. : 1991), 22(5), 1118–23. doi:10.1093/cercor/bhr186
Ung, H., Brown, J. E., Johnson, K. a, Younger, J., Hush, J., & Mackey, S. (2012). Multivariate Classification of Structural MRI Data Detects Chronic Low Back Pain. Cerebral cortex (New York, N.Y. : 1991), 1–8. doi:10.1093/cercor/bhs378
Wise, R. G., Rogers, R., Painter, D., Bantick, S., Ploghaus, A., Williams, P., Rapeport, G., et al. (2002). Combining fMRI with a Pharmacokinetic Model to Determine Which Brain Areas Activated by Painful Stimulation Are Specifically Modulated by Remifentanil. NeuroImage, 16(4), 999–1014. doi:10.1006/nimg.2002.1146
Wise, R. G., Williams, P., & Tracey, I. (2004). Using fMRI to quantify the time dependence of remifentanil analgesia in the human brain. Neuropsychopharmacology : official publication of the American College of Neuropsychopharmacology, 29(3), 626–35. doi:10.1038/sj.npp.1300364
Yarkoni, T., Poldrack, R. A., Nichols, T. E., Van Essen, D. C., & Wager, T. D. (2011). Large-scale automated synthesis of human functional neuroimaging data. Nature methods, 8(8), 665–674. doi:10.1038/nMeth.1635
Zhang, Z. G., Hu, L., Hung, Y. S., Mouraux, a, & Iannetti, G. D. (2012). Gamma-band oscillations in the primary somatosensory cortex--a direct and obligatory correlate of subjective pain intensity. The Journal of neuroscience : the official journal of the Society for Neuroscience, 32(22), 7429–38. doi:10.1523/JNEUROSCI.5877-11.2012