The Functional Anatomy of Nociception: Effective Connectivity in Chronic Pain and Placebo Response

Discussion

The most striking result of our study includes the following findings: recurrent effective connectivity within the lateral frontal pole becomes more inhibitory, while backward effective connectivity (from higher to lower cortical regions) decreases in both pain perceivers (as opposed to non-perceivers) and placebo responders (compared to non-responders). However, opposite changes are observed in forward connections, where nociception is associated with more excitatory (disinhibited) connections, while placebo responses evince more inhibitory forward connections. The changes in effective connectivity among pain perceivers were only observed in the left hemisphere. In leave-one-out cross-validation analyses, we found that the top-down connection from the left FP1 to the left SSC exhibited a sufficiently large (out-of-sample) effect size to predict whether an individual is experiencing knee pain or not. In a similar analysis of placebo response, we observed that the self-connection of the right insular demonstrated a sufficiently large effect size to predict placebo responses.

There are a few neuroimaging studies that have explored changes in functional and effective connectivity within nociceptive brain regions as potential biomarkers for chronic pain and placebo response (Liao et al., 2010; Liu et al., 2015; Cui et al., 2016; Tadayonnejad et al., 2016; Li et al., 2019; Lee et al., 2020). However, our research takes a unique approach, driven by a novel perspective on the neural mechanisms underlying sensory perception. There is a growing consensus that perception is not simply a passive process of progressively abstracting sensory input in a “bottom-up” manner. Instead, it involves both forward and backward information flow between brain regions organized in a hierarchical fashion, which plays a pivotal role in shaping perception. This concept forms the foundation of much of the current thinking about the functional architecture of the brain. One prominent theory in this realm is predictive coding (Mumford, 1992; Rao and Ballard, 1999; Friston and Kiebel, 2009), which has also been extended to the domain of motor function (Friston et al., 2011; Adams et al., 2013). The implicit exchange of neuronal messages within cortical hierarchies motivated are characterization of effective connectivity among pain processing regions organized hierarchically, and its potential relationship with pain perception and placebo response.

The modulation of effective connectivity in osteoarthritic patients and in placebo responders has some commonalities (for example, similar changes in backward and recurrent connections in the highest node, i.e., lateral pole) and some divergence (for example, opposite changes in forward connections). These changes are consistent with predictive coding accounts of pain perception and placebo response. For example, our study found that, in patients with osteoarthritic pain, top-down connections become more inhibitory and bottom-up connections more excitatory in the network involving interoceptive and somatosensory regions. These differences are consistent with the role of top-down predictions explaining away prediction errors at lower levels, as proposed by the predictive coding framework (Friston and Kiebel, 2009; Friston, 2012).

In particular, they are what would be predicted in terms of hierarchical predictive coding in which precision weighted prediction errors are passed forward to deeper levels of the interoceptive hierarchy to update or revise representations at higher levels (Clark, 2013; Gu et al., 2015b; Seymour and Mancini, 2020; Friston, 2023). An increase in forward connectivity can be read as an increase in the sensitivity of higher levels to ascending prediction errors. This corresponds to an increase in the precision of ascending prediction errors in people who experience pain or, conversely, an effective decrease in the precision of nociceptive prediction errors in people who show a placebo response. We will return to the important notion of precision and its neurophysiological correlates below.

Although long-range connections in the brain are excitatory (i.e., glutamatergic), predictive coding proposes that backward connections may preferentially target inhibitory interneurons in superficial and deep layers to evince an overall decrease in neuronal message passing (Friston and Kiebel, 2009; Friston, 2012; Aitchison and Lengyel, 2017). In predictive coding, this is often read as “explaining away” prediction errors at lower levels in sensory cortical hierarchies so that only those incoming stimuli that deviate from prediction (i.e., prediction errors) ascend the hierarchy to revise presentations at higher levels (Friston, 2012; Aitchison and Lengyel, 2017; Ray et al., 2020). However, top-down predictions also predict the reliability or precision of prediction errors at lower levels, leading to a disinhibitory modulation of lower-level activity in populations encoding prediction errors: sometimes discussed in terms of attention (Kok et al., 2012) or retrieval (Barron et al., 2020). One can associate this effect of top-down modulation with the group mean positive modulatory effects reported above. Thus, effective connectivity in chronic pain patients appear to reflect an enhanced pain processing within the nociceptive pathway with an increase in forward connectivity corresponding to an increase in the effects of ascending or bottom-up prediction error signaling.

An intriguing finding in connectivity changes in chronic pain patients—that warrants further comment—is that differences in connectivity are restricted within the left hemisphere, without any notable changes observed in the right hemisphere. However, note that in nearly all of the patients analyzed in the present study, OA was localized to the right knee, with only a handful experiencing bilateral knee involvement. The observation aligns with the anatomy of second-order pain neurons crossing over to the opposite side of the spinal cord and thus affording a potential explanation for left lateralized changes in connectivity. However, given prior evidence of lateralized pain processing in the brain (Coghill et al., 2001; Lu et al., 2004; Symonds et al., 2006), further research is needed to determine whether this effect is purely somatotopic or indicative of broader lateralization in pain modulation.

Current formulations of nociception—and in particular, the placebo effect-rest upon predictive coding and active inference accounts of hierarchical processing within the somatosensory and interoceptive hierarchy (Kube et al., 2020; Seymour and Mancini, 2020; Gu et al., 2015a). In particular, there is a focus on nuancing the perception of pain by adjusting the confidence or precision associated with the implicit (Bayesian) belief updating (Hoskin et al., 2019; Kube et al., 2020; Arandia and Di Paolo, 2021; Milde et al., 2023; Pagnini et al., 2023).

In brief, it may be the case that placebo effects can be attributed to a decrease in sensory precision of the kind associated with sensory attenuation (Limanowski, 2017; Wiese, 2017; Kube et al., 2020). Or, equivalently, an increase in the (subpersonal) confidence or precision afforded prior beliefs induced by the administration of placebos. Both or either of these changes in precision will move posterior beliefs toward a prior expectation that “I am not in pain because I have taken an analgesic”. In terms of predictive coding, this would correspond to an increased gain or precision weighting of ascending somatosensory and nociceptive prediction errors, relative to the precision of prior beliefs deeper in the interoceptive hierarchy (e.g., anterior insular and other prefrontal regions). From a psychological perspective, increases and decreases in the likelihood or sensory precision can be associated with selective attention or sensory attenuation, respectively. Physiologically, this kind of top-down precision weighting is thought to be mediated by selective changes in the postsynaptic sensitivity of certain neuronal populations: e.g., superficial pyramidal cells encoding prediction errors (Bastos et al., 2012; Friston, 2023). It is precisely (sic) this modulation of synaptic excitability that is measured by effective connectivity and evident in our DCM results.

In short, the changes in effective connectivity are consistent with a predictive coding formulation, in the following sense. Increased bottom-up prediction error corresponds to heightened pain perception, simply because the ascending prediction errors have been afforded more precision and therefore have more influence on belief updating processes at higher levels of the hierarchy. Similarly, reduced bottom-up prediction error signaling combined with increased top-down predictions are characteristic of placebo responders, suggesting that an increase in the precision or synaptic gain at higher hierarchical levels mitigates the accumulation of weak evidence (i.e., imprecise nociceptive prediction errors) for the high level belief: “I am in pain”.

It is important to qualify the interpretation of the DCM results in terms of predictive coding and predictions of precision. To draw definitive conclusions about differences in the precision of ascending prediction errors, it would require a DCM whose functional form was isomorphic with predictive coding schemes; namely, neuronal architectures equipped with precision or gain control of the sort used in computational neuroscience (Feldman and Friston, 2010; Brown and Friston, 2012; Moran et al., 2013; Pinotsis et al., 2014; Adams et al., 2015; FitzGerald et al., 2015; Parr et al., 2018). However, the requisite models would be too expressive to be identified or inverted using fMRI data. This means that dynamic causal models—with the requisite detail—are generally limited to informative EEG or MEG data. These (neural mass) models can support fine-grained inferences about changes in the excitability of superficial pyramidal cells, which are thought to encode uncertainty or precision. Please see Auksztulewicz and Friston (2015) and Pinotsis et al. (2014) for a discussion and empirical examples.

In dynamic causal models of fMRI, one often restricts the interpretation—in terms of precision or attentional gain—to intrinsic connections that determine the postsynaptic gain or excitability of neuronal populations. The synaptic mechanisms usually invoked rest upon fast-spiking inhibitory interneurons and modulatory neurotransmission (Shipp, 2016; Barron et al., 2020). When applying DCM to fMRI data, this places emphasis on the (inhibitory) self or recurrent connections that model the excitability of neuronal populations at each hierarchical level (via disinhibition). However, this coarse-grained modeling precludes any assertions about the neuronal populations involved or the synaptic mechanisms mediating the encoding of precision.

We would also like to offer a few further clarifications in response to a reviewer’s comment. First, our analyses focus on the modulation of effective connectivity associated with pain perception and placebo response, rather than mean resting-state connectivity, which is influenced by processes unrelated to pain and further confounded by averaging across heterogeneous groups. Second, in our findings, top-down connections became more negative with both pain perception and placebo response, a pattern consistent with increased prediction and/or reduced precision. While this may appear paradoxical, it can be understood in terms of the predictive content: chronic pain patients may rigidly apply strong pain predictions (“I am in pain”) even when sensory evidence contradicts them, due to low precision weighting, thereby sustaining pain perception. Conversely, placebo responders may hold strong predictions of relief (“This pill will help”) that override pain signals, also facilitated by reduced precision. Both scenarios thus reflect hyperpredictive states shaped by low precision, differing primarily in the content of the prediction. That said, we do not claim that these findings provide definitive evidence for a predictive coding framework, but rather that they are consistent with such an account while remaining interpretable independently of it.

The model comparison discussed above furnishes clear evidence for changes in a number of connections that underwrite nociception and placebo response. One might ask whether these changes can be used diagnostically in individual patients. In other words, are the underlying effect sizes sufficiently large to predict whether somebody is a patient or a control? Or anticipate the placebo response among patients? This question goes beyond whether there is evidence for an association and addresses the utility of connectivity phenotyping for precision medicine. One can estimate out of sample effect sizes using cross-validation under parametric empirical Bayesian schemes (Friston et al., 2016a). In this analysis, we withheld a particular participant and asked whether one could have predicted the group membership, given the effective connectivity estimates from that subject. In the current analysis, every connection showed a significant out-of-sample correlation with group membership for patient versus control and placebo responders versus non-responders analysis. This suggests that a nontrivial amount of variance in the group membership could be explained by effective connectivity.

Details of the leave-one-out cross-validation procedure can be found in the following tutorial papers (Zeidman et al., 2019a, 2019b).

As noted by one of our reviewers, the model comparison—using BMR and averaging—and cross-validation—using a leave-one-out procedure—highlighted differences in effective connectivity that were consistent but not identical. More specifically, the cross-validation analysis demonstrated that (i) what is predictive of pain experience is the strength of the backward connection from FPL to SSC and (ii) what is predictive of placebo response is the self-inhibitory connection from PI to itself—as identified with Bayesian model comparison. However, Bayesian model comparison also provided evidence for differences in forward connectivity, which do not appear to have predictive validity.

This apparent discrepancy reflects the use of the Bayesian model to identify changes in effective connectivity—in terms of the evidence for those changes—while the cross-validation was used to address predictive validity. In more detail, hypothetical changes were evaluated in terms of their evidence; namely, a variational bound on the (log) probability of the group data under each hypothesis or model. This is distinct from the cross-validation analyses, which provide an out-of-sample estimate of the effect sizes. In other words, the cross-validation analyses ask a different question; namely, given a new subject, could estimates of their effective connectivity predict whether they were experiencing pain or whether they would respond to placebo? Usually, only differences in connectivity that have a large effect size feature this kind of predictive validity.

A note on our choice of network nodes: as we were primarily interested in quantifying top-down or backward and bottom-up or forward connectivity in the cortical hierarchy, we selected two primary sensory cortices from brain regions sensitive to pain perception and one of the highest regions in pain pathway. Thus, bilateral primary somatosensory cortices and PI were selected as lower nodes. It should be pointed out here that PI is widely considered as the primary interoceptive cortex (Nieuwenhuys and Oudejans, 2012; Barrett and Simmons, 2015; Wilson-Mendenhall et al., 2019). As higher regions, we chose the right and left lateral frontal pole. Several tracing, lesion, and physiological studies suggest that visual, auditory, and somatosensory processing pathways converge at different regions of ventrolateral prefrontal cortex (VL PFC; Romanski, 2012; Spitzer et al., 2014) and dorsolateral prefrontal cortex (DLPFC; Meienbrock et al., 2007; Zhao and Ku, 2018). We therefore chose the lateral frontal pole as representative of a higher node. Empirical studies (Badre, 2008; Dumontheil, 2014) support a posterior to anterior sensory representational hierarchy in the prefrontal cortex and place the lateral frontal pole one level higher than both DL and VL PFC in the cortical hierarchy. Involvement of lateral frontal pole in pain perception is well established in several neuroimaging and magnetic stimulation studies (Feitosa et al., 2020; Ushio et al., 2020; Smith et al., 2021). Thus, we examined changes in the overall top-down and bottom-up effective connectivity—with pain perception and placebo response—by selecting nodes at both the highest and lowest levels of the cortical hierarchy in the pain processing pathway.

The findings from the current study should be interpreted in light of certain limitations. First, our study focused on patients with knee OA, a well-known cause of chronic pain. To ascertain whether the observed changes in connectivity represents a general pattern associated with chronic pain or a specific pattern linked to knee OA pain, it is essential to replicate this analysis in other chronic pain conditions. Second, when considering our connectivity analysis, it is crucial to acknowledge the potential presence of confounding factors beyond the age and sex of the participants. For instance, depression and anxiety frequently co-occur with chronic pain and may influence top-down effective connectivity in the brain. While none of our participants reported a diagnosis of major psychiatric conditions, we did not specifically rule out, or control for, the presence of subclinical depression or anxiety.

The findings from this study hold considerable promise for practical applications. Future research could be aimed at assessing the efficacy of therapeutic interventions, encompassing various pharmacological and non-pharmacological treatments, in reversing the alterations in cortical effective connectivity and pain perception. An intriguing avenue to explore involves the use of emerging noninvasive brain stimulation techniques, such as transcranial magnetic stimulation (TMS). Recent studies have demonstrated the ability of TMS to modulate cortico-cortical connectivity within specific neural circuits (Fox et al., 2012; Groppa et al., 2013; Giambattistelli et al., 2014). By applying these techniques to target specific brain regions within the pain processing pathway, we can investigate their impact on nociception using state-of-the-art methodologies that are currently available.