Periodic Reporting for period 4 - PredictingPain (Deconstructing pain with predictive models: from neural architecture to pain relief)
Reporting period: 2023-04-01 to 2024-09-30
Pain is a useful and important warning signal of impending or actual tissue damage, but in its chronic form it is also a troubling healthcare issue. Chronic pain is a burden for millions of people in terms of suffering, as well as for societies in terms of costs. In Europe, about 20% of adults suffer from chronic pain, resulting in tremendous societal impact in terms of increased use of healthcare resources, decreased quality of life, and decreased ability to participate in the workforce. Thus, there is an obvious need for improved treatment that in turn calls for a better understanding of the neurobiological processes underlying pain.
It is important to point out that in many circumstances the subjectively perceived intensity of pain is not necessarily related to the amount of peripherally received nociceptive input. This suggests a strong involvement of the central nervous system (CNS) in the construction of pain, however a unifying mechanistic framework for how CNS processing shapes pain perception is still lacking. Here we aim to provide such a framework by taking a novel approach to pain that capitalizes on recent advances in neuroscience that characterize perception not as a passively arising response to sensory stimuli, but as an active process, in which the CNS constantly generates predictions about the sensory inputs it receives and adjusts these predictions in light of new sensory input.
One implementation of this general idea is termed predictive coding and has been applied especially in the visual and auditory domain over the last years, where it was able to explain numerous perceptual and neurobiological phenomena. Predictive coding rests on the idea that each area in a sensory hierarchy contains two distinct populations of neurons: ‘prediction units’ that represent the currently chosen hypothesis of sensory input and ‘prediction error units’ that represent the mismatch between the sensory input and the current hypothesis. In opposition to traditional models, which assume that sensory input is passed forward along the sensory hierarchy, predictive coding assumes that only the unexplained part of sensory input is passed forward (via prediction error units) and that it is the prediction units that contain the current hypothesis of sensory input. Together, these units are supposed to be organized in a hierarchical architecture, where prediction error units carry feedforward activity and prediction units carry feedback activity, with the goal of optimizing predictions and reducing prediction errors.
Interestingly, this framework has not yet received a large amount of attention in the context of pain or nociceptive processing and one might also question whether neurobiological processes in the nociceptive system – or the subjective experience of pain – are suitable subjects for this approach, since they are rather removed (both in terms of anatomy and experience) from the visual system, where the foundations of predictive coding were established. In this project, we will therefore investigate if basic principles of predictive coding also apply in the context of pain. We aim to investigate predictive coding mechanisms at all levels of the CNS, but will especially focus on the human spinal cord, using advanced neuroimaging techniques that should allow for novel insights into this lowest and earliest level of the nociceptive processing hierarchy. This is motivated by the dorsal horn of the spinal cord being the first station of CNS pain processing, and also a structure that is involved in pathological pain processes. Most importantly though, the spinal cord is of strategic importance, as predictive signals at this lowest level of the CNS might have a profound effect on feedforward propagation of prediction errors to higher levels and thus the ensuing pain perception.
While not all studies of this project have been finished, there are several conclusions that can already be drawn. First, we did not obtain consistent evidence for an influence of predictive processes across the studies carried out, but observed a more nuanced picture. For example, while we were able to observe that participants in our studies did indeed form expectations (or generate predictions) about forthcoming pain, these did not influence sensory processing at all levels: neither highly synchronized pain-induced neuronal signals in cortex (as for example measured via electroencephalography) nor participants’ perceptual performance in pain-discrimination tasks were influenced by predictions. Also, previously observed prediction error signals were not a consistent feature in our results, even though we had optimized our experiments towards their detection. At least from this limited set of experiments carried out, we can conclude that predictive coding principles might not consistently apply to all aspects of pain perception and associated processing.
Second, we were able to achieve several important methodological advances with regards to advanced neuroimaging of the human spinal cord. On the one hand, we developed a non-invasive electrophysiological method for the recording of spinal cord signals with high sensitivity and temporal precision. This method also allowed for initial insights at the spinal level regarding integrative processes and early nociceptive processing. On the other hand, we harnessed the benefits of advanced magnetic resonance imaging methods and developed an approach that allowed for investigating pain-induced spinal cord responses with very high spatial resolution – here, we have been able to map different response components onto different fine-grained anatomical compartments of the spinal cord.
Third, in ongoing work arising from this ERC project we are investigating whether we can uncover an influence of predictive processes on nociceptive processing at the spinal, now capitalizing on the heightened sensitivity and resolution provided by the newly developed methods.
It is important to point out that in many circumstances the subjectively perceived intensity of pain is not necessarily related to the amount of peripherally received nociceptive input. This suggests a strong involvement of the central nervous system (CNS) in the construction of pain, however a unifying mechanistic framework for how CNS processing shapes pain perception is still lacking. Here we aim to provide such a framework by taking a novel approach to pain that capitalizes on recent advances in neuroscience that characterize perception not as a passively arising response to sensory stimuli, but as an active process, in which the CNS constantly generates predictions about the sensory inputs it receives and adjusts these predictions in light of new sensory input.
One implementation of this general idea is termed predictive coding and has been applied especially in the visual and auditory domain over the last years, where it was able to explain numerous perceptual and neurobiological phenomena. Predictive coding rests on the idea that each area in a sensory hierarchy contains two distinct populations of neurons: ‘prediction units’ that represent the currently chosen hypothesis of sensory input and ‘prediction error units’ that represent the mismatch between the sensory input and the current hypothesis. In opposition to traditional models, which assume that sensory input is passed forward along the sensory hierarchy, predictive coding assumes that only the unexplained part of sensory input is passed forward (via prediction error units) and that it is the prediction units that contain the current hypothesis of sensory input. Together, these units are supposed to be organized in a hierarchical architecture, where prediction error units carry feedforward activity and prediction units carry feedback activity, with the goal of optimizing predictions and reducing prediction errors.
Interestingly, this framework has not yet received a large amount of attention in the context of pain or nociceptive processing and one might also question whether neurobiological processes in the nociceptive system – or the subjective experience of pain – are suitable subjects for this approach, since they are rather removed (both in terms of anatomy and experience) from the visual system, where the foundations of predictive coding were established. In this project, we will therefore investigate if basic principles of predictive coding also apply in the context of pain. We aim to investigate predictive coding mechanisms at all levels of the CNS, but will especially focus on the human spinal cord, using advanced neuroimaging techniques that should allow for novel insights into this lowest and earliest level of the nociceptive processing hierarchy. This is motivated by the dorsal horn of the spinal cord being the first station of CNS pain processing, and also a structure that is involved in pathological pain processes. Most importantly though, the spinal cord is of strategic importance, as predictive signals at this lowest level of the CNS might have a profound effect on feedforward propagation of prediction errors to higher levels and thus the ensuing pain perception.
While not all studies of this project have been finished, there are several conclusions that can already be drawn. First, we did not obtain consistent evidence for an influence of predictive processes across the studies carried out, but observed a more nuanced picture. For example, while we were able to observe that participants in our studies did indeed form expectations (or generate predictions) about forthcoming pain, these did not influence sensory processing at all levels: neither highly synchronized pain-induced neuronal signals in cortex (as for example measured via electroencephalography) nor participants’ perceptual performance in pain-discrimination tasks were influenced by predictions. Also, previously observed prediction error signals were not a consistent feature in our results, even though we had optimized our experiments towards their detection. At least from this limited set of experiments carried out, we can conclude that predictive coding principles might not consistently apply to all aspects of pain perception and associated processing.
Second, we were able to achieve several important methodological advances with regards to advanced neuroimaging of the human spinal cord. On the one hand, we developed a non-invasive electrophysiological method for the recording of spinal cord signals with high sensitivity and temporal precision. This method also allowed for initial insights at the spinal level regarding integrative processes and early nociceptive processing. On the other hand, we harnessed the benefits of advanced magnetic resonance imaging methods and developed an approach that allowed for investigating pain-induced spinal cord responses with very high spatial resolution – here, we have been able to map different response components onto different fine-grained anatomical compartments of the spinal cord.
Third, in ongoing work arising from this ERC project we are investigating whether we can uncover an influence of predictive processes on nociceptive processing at the spinal, now capitalizing on the heightened sensitivity and resolution provided by the newly developed methods.
We have developed several experimental paradigms allowing us to investigate the relevance of predictions (or, as they are often also referred to: expectations) and the underlying neural signals from different perspectives. This is done for example by experimentally manipulating i) the temporal structure of pain stimulation (e.g. one pain stimulus predicting the occurrence of another one), ii) the probability by which painful stimuli will occur in a certain context or iii) the location at which painful stimuli will most likely occur.
With regards to the first example (temporal structure of painful stimulation), we have replicated and extended the known finding that pain-induced cortical electroencephalography (EEG) responses are reduced for a temporally-predictable second painful stimulus compared to a first painful stimulus. Importantly however, the degree to which this occurs was not dependent of the expectations that participants formed about this relationship. Second, regarding the association between a context and the probability of receiving pain, we were able to demonstrate that contexts with a high likelihood of pain delivery resulted in a more intense experience of pain, despite no change in the actual stimulation. The expectations that participants generated in these contexts were also evident in metrics that index autonomic nervous system activation, such as changes in sweating (skin conductance responses) and pupil size (pupil dilation responses): higher probabilities of receiving pain resulted in higher skin conductance and higher pupil dilation responses to pain, despite the physical stimulus intensity delivered to the skin of participants being identical across these conditions. Interestingly, while we were thus able to clearly demonstrate the behavioural coding of expectation/prediction signals, we were not able to obtain any evidence for a behavioural representation of prediction errors, i.e. responses that are driven by a mismatch between the predicted and the actual occurrence of painful stimulation. Third, providing cues that signaled to which body part painful stimuli would soon be delivered generated anticipatory skin conductance and pupil dilation changes, with the latter also showing evidence of prediction errors when spatial expectations were violated. Crucially however, the expectations generated in these situations did not influence participants perceptual performance in a pain discrimination task at the predicted body location.
In addition, we have carried out several methodological projects, for example to improve functional magnetic resonance imaging (fMRI) data quality from the human spinal cord, both at the field strength of 3 Tesla (which is the current standard for spinal cord fMRI) and at the field strength of 7 Tesla (which is needed in order to obtain high-resolution fMRI data). For example, at 3 Tesla we have devised a user-friendly method that allows for an automated calculation of so-called ‘shim-settings’ which make the magnetic field more homogenous and contribute to better data quality. At 7 Tesla, we have developed approaches for obtaining high-resolution fMRI data, allowing us non-invasively investigate spinal processing of nociceptive signals at a previously not available spatial resolution. Complementing these hemodynamic approaches, we have developed a kind of ‘spinal cord EEG’, whereby we record multi-channel electrophysiological data from the spinal cord via a large number surface electrodes placed on participants’ neck and back. All of these approaches have allowed more precise insights into bottom-up processing of nociceptive information at the spinal level and with ongoing work we are aiming to extend this to characterizing top-down signals as needed for a fundamental assessment of predictive coding principles.
In terms of exploitation and dissemination, several aspects are worth mentioning. First, all data generated in this project have been (and are being) made openly available upon publication of the respective manuscripts and can thus be re-used. Second, all analysis code project is shared via public repositories, including novel methods for data acquisition, thus facilitating uptake by other researchers. Third, some of the methods developed in this project have been contributed to and incorporated in open-source software tools for neuroimaging data analysis. Fourth, manuscripts are first shared as preprints, are then published in gold open-access format and are accompanied by press releases in order to maximize outreach. Finally, the group has not only presented the results that have arisen from this project at various international conferences and workshops but has also been active in outreach activities (also targeting various age groups, i.e. from school-age children to senior citizens).
With regards to the first example (temporal structure of painful stimulation), we have replicated and extended the known finding that pain-induced cortical electroencephalography (EEG) responses are reduced for a temporally-predictable second painful stimulus compared to a first painful stimulus. Importantly however, the degree to which this occurs was not dependent of the expectations that participants formed about this relationship. Second, regarding the association between a context and the probability of receiving pain, we were able to demonstrate that contexts with a high likelihood of pain delivery resulted in a more intense experience of pain, despite no change in the actual stimulation. The expectations that participants generated in these contexts were also evident in metrics that index autonomic nervous system activation, such as changes in sweating (skin conductance responses) and pupil size (pupil dilation responses): higher probabilities of receiving pain resulted in higher skin conductance and higher pupil dilation responses to pain, despite the physical stimulus intensity delivered to the skin of participants being identical across these conditions. Interestingly, while we were thus able to clearly demonstrate the behavioural coding of expectation/prediction signals, we were not able to obtain any evidence for a behavioural representation of prediction errors, i.e. responses that are driven by a mismatch between the predicted and the actual occurrence of painful stimulation. Third, providing cues that signaled to which body part painful stimuli would soon be delivered generated anticipatory skin conductance and pupil dilation changes, with the latter also showing evidence of prediction errors when spatial expectations were violated. Crucially however, the expectations generated in these situations did not influence participants perceptual performance in a pain discrimination task at the predicted body location.
In addition, we have carried out several methodological projects, for example to improve functional magnetic resonance imaging (fMRI) data quality from the human spinal cord, both at the field strength of 3 Tesla (which is the current standard for spinal cord fMRI) and at the field strength of 7 Tesla (which is needed in order to obtain high-resolution fMRI data). For example, at 3 Tesla we have devised a user-friendly method that allows for an automated calculation of so-called ‘shim-settings’ which make the magnetic field more homogenous and contribute to better data quality. At 7 Tesla, we have developed approaches for obtaining high-resolution fMRI data, allowing us non-invasively investigate spinal processing of nociceptive signals at a previously not available spatial resolution. Complementing these hemodynamic approaches, we have developed a kind of ‘spinal cord EEG’, whereby we record multi-channel electrophysiological data from the spinal cord via a large number surface electrodes placed on participants’ neck and back. All of these approaches have allowed more precise insights into bottom-up processing of nociceptive information at the spinal level and with ongoing work we are aiming to extend this to characterizing top-down signals as needed for a fundamental assessment of predictive coding principles.
In terms of exploitation and dissemination, several aspects are worth mentioning. First, all data generated in this project have been (and are being) made openly available upon publication of the respective manuscripts and can thus be re-used. Second, all analysis code project is shared via public repositories, including novel methods for data acquisition, thus facilitating uptake by other researchers. Third, some of the methods developed in this project have been contributed to and incorporated in open-source software tools for neuroimaging data analysis. Fourth, manuscripts are first shared as preprints, are then published in gold open-access format and are accompanied by press releases in order to maximize outreach. Finally, the group has not only presented the results that have arisen from this project at various international conferences and workshops but has also been active in outreach activities (also targeting various age groups, i.e. from school-age children to senior citizens).
First, we believe that the ultra-high-field fMRI approach we have developed in order to non-invasively investigate pain-related responses in the human spinal cord is currently unique in terms of its ability to resolve fine-grained responses: we are using fMRI with a spatial resolution of 750 x 750 micrometres. When relating this to known anatomical data of spinal cord organization, it becomes clear that with this resolution we are in a spatial range that allows for differentiating the functionally relevant units: for example, we have mapped two different bottom-up response components onto deep versus superficial layers of the spinal cord.
Second, the multi-channel electrophysiological approach for non-invasive spinal cord recordings developed here offers a novel and unique window into human spinal cord function. While it comes with single-trial sensitivity and a temporal resolution in the sub-millisecond range, its spatial resolution is rather low and this method thus complements the fMRI approach mentioned above. We have employed and are currently using this approach to investigate various bottom-up aspects of nociceptive processing with simultaneous spinal and cortical recordings, with future work focused on interactions between these levels.
Overall, we focused on providing a characterization of pain processing at the level of the human spinal cord – an area that, compared to the brain, has received less attention in human pain research and was the main anatomical target of this project. Due to unanticipated technical and logistical difficulties, as well as the fact that prediction effects were not observed to be as robust as anticipated in several of our experiments, we however had to put a stronger focus on bottom-up processing compared to predictive top-down processing than initially planned in the project. Nevertheless, we expect that we will be able to make use of the above-mentioned methodological developments to provide an answer to the question of whether fundamental properties of the predictive coding framework also apply in the domain of pain at the level of spinal cord processing.
Second, the multi-channel electrophysiological approach for non-invasive spinal cord recordings developed here offers a novel and unique window into human spinal cord function. While it comes with single-trial sensitivity and a temporal resolution in the sub-millisecond range, its spatial resolution is rather low and this method thus complements the fMRI approach mentioned above. We have employed and are currently using this approach to investigate various bottom-up aspects of nociceptive processing with simultaneous spinal and cortical recordings, with future work focused on interactions between these levels.
Overall, we focused on providing a characterization of pain processing at the level of the human spinal cord – an area that, compared to the brain, has received less attention in human pain research and was the main anatomical target of this project. Due to unanticipated technical and logistical difficulties, as well as the fact that prediction effects were not observed to be as robust as anticipated in several of our experiments, we however had to put a stronger focus on bottom-up processing compared to predictive top-down processing than initially planned in the project. Nevertheless, we expect that we will be able to make use of the above-mentioned methodological developments to provide an answer to the question of whether fundamental properties of the predictive coding framework also apply in the domain of pain at the level of spinal cord processing.