Multiple Functional Brain Networks Related to Pain Perception Revealed by fMRI

33 healthy, right-handed adults (22 females, 11 males) participated in the study, with a mean age of 27.9 years (SD = 9.0 years). All participants provided informed consent and reported no prior history of psychiatric, neurological or pain disorders. Ethical review and approval were provided by the Columbia University Institutional Review Board (Protocol number AAAE3743). Since the data were anonymized and we performed a secondary analysis, no local ethics review was required.

In our study, two participants (subjects 11 and 30) were excluded because they received too few trials under each experimental condition (defined below), creating problems for the fMRI-CPCA algorithm. This left 31 participants to be analysed.

To elicit pain, a thermode device was placed on the volar surface of the left forearm (TSA-II Neurosensory Analyzer with a 16-mm Peltier thermode endplate, Medoc Advanced Medical Systems). Thermal stimuli were delivered at specific temperatures for 12.5 s each, with 3 s of ramp-up, 7.5 s at the target temperature, and 2 s of ramp-down. Temperature levels ranged from 40.8 °C to 47.3 °C (study documentation and participant results are available at https://​openneuro.​org/​datasets/​ds000140/​versions/​00001).

Participants completed 9 separate functional scanning sessions. There were 3 types of sessions: “standard” runs, where pain stimulation was received passively; “regulate-up” runs, where participants were instructed to increase the intensity of pain by cognitive control; and their counterpart, “regulate-down” runs. The regulation manipulation was intended to engage supplementary brain systems for pain regulation. For the explicit purposes of our study, we focused on standard runs only.

Each standard run began with an 18-s fixation cross presented on screen, followed by 11 consecutive trials. Each trial was 33–41 s long and featured the same progression: 12.5 s of thermal stimulation, 4.5–8.5 s (jittered) of pre-rate rest, 11 s of pain rating (completed on screen using a hand-held remote), and 5–9 s (jittered) of post-rate rest. The rating period involved two kinds of rating; first, participants decided whether a stimulus was painful or not (this phase lasted 4 s), then participants rated the intensity of their sensation on a Visual Analogue Scale from 0 to 200, where the interval 0–100 represented non-painful warmth, and 100–200 represented the intensity of a stimulus perceived as painful. The scale was presented on screen, and participants were instructed as to the meaning of each interval prior to scanning. The specific order of temperatures administered throughout each run can be found in Woo et al. (2015). For a schematic illustration of task design, see Fig. 1.

Whole-brain functional images were collected on a 3 T Philips Achieva TX scanner at Columbia University’s Program for Imaging in Cognitive Science (PICS). Structural images were collected with high-resolution T1 spoiled gradient recall images (SPGR), which allow for anatomical localization and warping to standard space. For functional EPI image collection, the following scanning parameters were set: TR  =  2000 ms, TE  =  20 ms, field of view  =  224 mm, 64 × 64 matrix, 3 × 3 × 3 mm³ voxels, 42 interleaved slices, parallel imaging, SENSE factor 1.5. E-Prime software (PST Inc.) was used to control stimulus presentation and collect behavioural data.

For our analysis, all preprocessing was completed in SPM 12. Structural and functional scans were reoriented manually, such that the origin was placed on the anterior commissure, and the AC-PC plane was oriented horizontally. Slice-time correction was performed to mitigate the temporal lag in slice acquisition across the 2-s TR, using slice 21 as a reference. Realignment algorithms were applied to counteract displacement of voxels due to head movement, and runs that exceeded movement parameter thresholds of 4.5 mm in either z, x, y direction, as well as pitch, yaw or roll, by at least 50 scans, were removed from the analysis (subject 10, run 5 and 6; subject 2, run 5; subject 4, run 1, 4 and 5). For each participant, functional scans were co-registered to their corresponding structural images, and structural T1 scans were segmented into gray matter, white matter, cerebrospinal fluid, meninges and skull components. Finally, raw functional data were normalized to MNI template space (with a voxel size of 3 × 3 × 3 mm) and smoothed with a 6 × 6 × 6 FWHM Gaussian kernel.

For detailed explanations of preprocessing methods, along with specific versions of software tools used, refer to supplementary materials.

Constrained Principal Component Analysis (CPCA) is a statistical technique that combines multivariate least-squares regression with principal component analysis (Hunter & Takane, 2002; Takane & Shibayama, 1991; Takane & Hunter, 2001). It can be used to perform whole-brain analyses of fMRI BOLD signal data. When applied to fMRI, it identifies multiple functional networks involved in a task and estimates fluctuations in BOLD signal for each network, over a specified interval of time. Further statistical tests can be used to quantify the interactions between networks, correlational relationships between network activation and behavioural measures, and the effect of experimental manipulations on the activation of each network.

Broadly speaking, fMRI-CPCA involves two steps. First, multivariate least-squares multiple regression is used to isolate variance in BOLD signal that is predictable from the timing of stimulus presentation, after which the variance is said to be “constrained” to task timing. This first step is referred to as the external analysis. Second, a principal component analysis (PCA) is conducted on the constrained portion of the variance in BOLD signal, and the extracted components represent systems of functionally interconnected voxels (i.e. functional brain networks) related to the task. This step is referred as the internal analysis. Importantly, applying PCA after the regression ensures that the networks identified are based on task-related information only. This is a defining feature of fMRI-CPCA and distinguishes it from other applications of PCA (or ICA) used in fMRI. In fMRI-CPCA, the variance shared between principal components and task timing is maximized, thus avoiding any contamination of the solution by variability that is not predictable from task timing. Ultimately, fMRI-CPCA outputs brain activity maps that can be overlaid on a structural image (for example, in applications like MRIcron [https://​www.​nitrc.​org/​projects/​mricron]), as well as estimated hemodynamic response shapes (plotted over post-stimulus time) for each combination of network, subject and task condition. The next few paragraphs will elaborate on specific matrices and equations required to implement the analyses.

In order to perform the external analysis, two matrices must first be prepared. The Z matrix (or activation matrix) contains the BOLD data for all runs, with each voxel represented as a single column, and each full-brain scan represented as a single row. In the current study, 31 subjects went through nine runs each, with 209 scans per run. Six runs were removed due to excessive head movement (see section “2.5. Preprocessing”), leaving a total of 42,427 rows (full brain scans) and 79,522 columns (voxels) in the Z matrix. The mean value for each voxel was centered to zero for each run separately, and the variables standardized (such that the standard deviations were set to one for each run separately). The G matrix (or design matrix) contains a Finite Impulse Response (FIR) model of the BOLD signal based on stimulus presentation timing; unlike more conventional models, the FIR model does not impose a predetermined HDR shape on the dataset (which is commonly assumed to aid in determining task-relevant activations in BOLD signal). Instead, a value of one is placed into cells of G for which the BOLD signal is to be estimated, and a value of zero is in all other cells—thus, the G matrix simply defines the time intervals during which we expect to see task-relevant activations. The number of rows in the G matrix will equal the number of rows in the Z matrix, but the number of columns is equal to the number of post-stimulus time points (time bins) for which the BOLD signal is to be predicted, multiplied by the total number of conditions and the total number of subjects. The G matrix is also standardized for each individual run. We then regress the Z matrix onto the G matrix, where C = (G ′ G)⁻¹G ′ Z is a matrix of timepoint- and voxel-specific regression weights that satisfy the least-squares criterion. When C is applied to G it provides a matrix of BOLD signal values predicted from task-timing, _scans \( \hat{Z} \)_voxels or GC. E represents the residual signal (i.e. signal that is not predictable from task-timing), which is disregarded in the rest of the analysis. As an additional note, E can be further analysed exactly like GC; such an analysis would produce the dominant networks that are not predictable from task timing, which may be those engaged during off-task periods, or task on processes that span the whole series of trials but are not specifically elicited by the onset of tasks. This type of analysis was not carried out here and is beyond the scope of this paper.

The next stage of fMRI-CPCA is the internal analysis, which typically involves application of a principal component analysis (PCA) to the constrained, task-related signal (GC). This identifies correlated structure underlying the voxel data, grouping correlated voxels into components that represent functional brain networks. Importantly, these components will be optimized to be task-related, because GC contains task-related variance only. PCA is achieved through singular value decomposition of GC: where U is a matrix of left singular vectors, D is a diagonal matrix of singular values, and V′ is a matrix of right singular vectors. In matrix U, columns represent components, and rows represent scans. The values in matrix U are “component scores” and provide an indication or “score” of how important each component is for each scan. In matrix V, columns represent components, and rows represent voxels. Cells of V can be rescaled by \( VD/\sqrt{N} \) to obtain “component loadings”—correlation coefficients indicating the correlation of task-related BOLD signal in each voxel with the respective component scores. Voxels that are highly correlated with a given component’s “component scores” form the brain regions that define the functional network represented by that component. Notably, rescaling right singular vectors in V allows them to be interpreted as correlations between voxels and networks, while also providing a better approximation of the inputted matrix GC by incorporating the variance accounted for by each network. To visualize a brain activity map for each network, columns in rescaled V were overlaid on a brain template in MRIcron and thresholded to display only the voxels with the most dominant loadings (e.g. top 10% absolute values). In the current study, we orthogonally rotated and rescaled the V matrix prior to display, using a varimax solution with 500 iterations (Abdi & Williams, 2010; Bryant & Yarnold, 1995; Kaiser, 1958).

PCA identifies a large number of components, but a select few can be extracted (the components that account for the least amount of variability are considered noise, or brain activity that is unlikely to be reliable). Various methods for component selection exist; in this study we used the elbow method. This method relies on visual inspection of the scree plot of singular values (Cattell, 1966; Cattell & Vogelmann, 1977). When plotted, singular values (which are contained in D) produce a line that gradually approaches zero as components account for less and less variance. In the elbow method, components are selected for extraction by locating the first abrupt increase in variance—relative to the baseline variance accounted for by the majority of components—and extracting the associated component followed by all components that account for a greater proportion of variance (Kodinariya & Makwana, 2013). In this study, 4 components were retained and varimax rotated. We attempted additional analyses with a greater number of components retained, to ensure the validity of the chosen threshold. Additional networks did not substantially improve the solution, only marginally increasing variance explained and failing to detect new and informative regions/networks, instead fragmenting networks previously identified in the four-component solution.

After the external and internal analyses are complete, a final step is applied to produce estimates of HDR shapes associated with each network. This is achieved by relating component scores (in matrix U) back to stimulus presentation timing (coded in G), and computing P such that: where P contains “predictor weights”—these are weights that estimate the intensity of each component for the time bins specified in G. When plotted over post-stimulus time, predictor weights reveal the unique HDR shapes elicited by each subject and condition within each network, for the specified interval of time. In this study, predictor weights were averaged over subjects before plotting. Further, predictor weights were averaged over post-stimulus time to compute overall intensity values for network activation; more detail on this is provided below.

The goal of the current study was to determine how the brain configures itself when processing pain, and to use the brain networks detected to generate a model of subjective pain perception. Accordingly, we formatted the G matrix such that separate HDR shapes would be produced for high and low temperature conditions. The division was based on the median temperature administered across all trials, including regulation runs (the median temperature was 44.3 °C, any stimulus that was equal to or less than 44.3 was assigned to the “low” temperature condition; the rest were “high”). We examined brain activity during thermal heat portions of the experimental task only, and only included standard runs in the analysis to avoid capturing brain systems for cognitive self-regulation over pain perception. The task-relevant time interval (encoded in G) was defined as the 16 s immediately following thermal stimulus presentation; in this way, the entire duration of the stimulus and 3.5 s thereafter were accounted for. Because each full-brain fMRI scan was completed in two seconds, HDRs were estimated for eight post-stimulus time bins. The G matrix therefore consisted of 496 columns (2 conditions × 31 subjects × 8 time bins), and 42,427 rows (equal to the number of rows in the Z matrix).

For each network, the predictor weights produced in the final step of fMRI-CPCA define the unique HDR shapes associated with each condition, over the specified 16-s time interval. To model pain perception from brain activation, it was preferable to compute a single value that would capture the intensity of the response. In this case, due to exploration of the HDR shapes obtained, network activation intensity was estimated by averaging predictor weights (i.e. estimated BOLD signal) over the entire post-stimulus time interval. This yielded 248 estimates in total, one for each temperature category for each of the four networks detected by fMRI-CPCA for every participant.

Pain ratings were subjected to a similar procedure: ratings associated with each temperature condition were averaged over trials to obtain participant-specific estimates of pain perception during high- and low-temperature stimuli, yielding 62 estimates in total.

Two separate multiple linear regression analyses were conducted on pain rating and network activation data.

The scree plot of singular values indicated that four components should be extracted. Components 1, 2, 3, and 4 accounted for 21.47, 7.26, 4.95, and 3.74% of task-related variance in BOLD signal, respectively. Component images and estimated HDR shapes are displayed in Figs. 2, 3, 4, 5, along with network activation intensities for each temperature category (box plots).

Component 1 was primarily comprised of a) motor areas, including the primary motor cortex (M1), supplementary motor area (SMA) and cerebellum, b) visual areas, including the lateral occipital cortex (LO), and c) the primary somatosensory cortex (S1). Component 2 featured a frontoparietal activity pattern that included activation peaks in the anterior cingulate cortex (ACC), dorsolateral prefrontal cortex (dlPFC), anterior and posterior insula (aIns and pIns), and thalamus. Component 3 was limited to the outer edge of the cortex (specifically the frontal and occipital poles) and the longitudinal fissure and was mostly composed of negative loadings (i.e. it became deactivated during stimulation). This particular configuration was biologically untenable and resembled no established networks; it was most likely summarizing head movement that was coordinated with the application of thermal pain. Component 4 was characterized by deactivations in areas conventionally associated with the default mode network, including the posterior cingulate cortex (PCC), medial prefrontal cortex (mPFC), precuneuous and angular gyrus (AnG). Detailed anatomical descriptions for all networks are found in supplementary information (supplementary tables 1–3).

To relate these networks back to pain perception, we modelled variability in pain ratings as a linear function of activation intensities in all networks, both within and across individuals. The dataset used to model pain ratings consisted of pain ratings and network activation intensities for each temperature category (i.e. high and low) within each participant. Descriptive statistics (means and standard errors) for these data are found in supplementary table 4.

Component 1, with prominent activation peaks in MI, SMA, and cerebellum, is most aptly described as a sensory and motor network. In the context of thermal stimulation, sensation and motor output may be related to an instinctive flexing or bracing, or a desire to move, in response to intense stimuli (Davis & Moayedi, 2013; Davis et al., 2002). Hemodynamic response shapes (HDRs) for Component 1 showed that activation was in fact exclusive to higher intensity stimuli (temperatures above 44.3 °C). Component 1 also included prominent activations in SI, which, as a key cortical aspect of the lateral nociceptive system, is one of the first recipients of ascending pain signals through the spinothalamic tract (Davis & Moayedi, 2013; Fomberstein et al., 2013; Vierck et al., 2013; Yam et al., 2018). SI’s inclusion in Component 1 suggests that, during thermal stimulation, motor processes are prioritized and closely coordinated with sensory-discriminative functions (e.g. determination of stimulus location and intensity). In theory, such close communication would be necessary for an effective pain avoidance response when stimulus intensity reaches noxious levels (Postorino et al., 2017). This plausible role of Component 1 in generating pain-induced motor commands remains to be further explored; follow-up studies would benefit from monitoring physical movements in conjunction with other variables, allowing for the precise relationships between Component 1 activation intensity, stimulus intensity (e.g. temperature), motion, and pain perception to be determined.

A novel observation from fMRI-CPCA is the temporal overlap between visual areas and sensory-motor coupling, evidenced in Component 1. Their detection is likely an idiosyncratic capture of fMRI-CPCA, which avoids using regions-of-interest to spatially constrain the analysis. In fact, the anatomy of Component 1 replicates previous applications of fMRI-CPCA in other domains—specifically, it resembles a network consistently associated with sensorimotor response processes, featuring activations in lateralized MI, SMA, SI, cerebellum, and visual areas including the lateral occipital cortex (LO; Goghari et al., 2017; Larivière et al., 2017; Metzak et al., 2011). In this case, sensorimotor-visual coupling was likely observed because of screen-related cues that coincided with stimulus presentation.

In agreement with previous studies, Component 2 incorporated a large number of regions involved in pain, and combined sensory, affective and cognitive sub-networks (Davis & Moayedi, 2013; Wilcox et al., 2015). For example, the most prominent activation peaks were found in SII and posterior insula (pIns; sensory-discriminative regions), dACC, aIns, and thalamus (affective-motivational regions), and dlPFC and IPL (cognitive-evaluative regions; Peyron et al., 1999; Wilcox et al., 2015). Based on this, Component 2 could reflect a unification of sensory, affective and cognitive processes (Melzack & Casey, 1968) into a coordinated pain response.

The blending of sub-networks is likely facilitated by their inter-connectivity at rest, provided by common nodes in ACC and aIns that serve as relay sites between sub-networks (Wilcox et al., 2015). Importantly, the ACC and aIns are engaged in non-specific “salience detection”, where stimuli are selected based on their behavioural relevance, and attentional systems are primed to enable an effective response (Legrain et al., 2011; Menon, 2015). Such a “salience network” receives axonal projections from sensory areas like the pIns, which are thought to provide the aIns with incoming sensory information (Menon & Uddin, 2010). The pattern of activation observed in Component 2 captures both attentional systems (i.e. the cognitive sub-network of pain) and sensory-discriminative elements like SII and pIns. Thus, Component 2 may represent a salience network-mediated response to salience—in this case thermal stimulation—or more precisely a sequential activation of sub-networks, i.e. sensory systems activate the salience network, which then activates cognitive systems for sustained attention. The directionality of sub-network relations is a matter of speculation, but it presents an interesting question for future investigation. Additionally, the putative attentional function of Component 2 may be further explored by analyzing its pain-induced response during experimental manipulations of attentional demand or stimulus salience; a larger effect of attention on network response than stimulus temperature would suggest an attentional role.

The tendency for brain areas to become deactivated during a task and engaged at rest gave rise to the original concept of the “default mode of brain function” (Shulman et al., 1997; Raichle et al., 2001). Since being originally characterized, research has emerged documenting the overall functional contributions of the default-mode network (DMN) to human behavior, including its relevance to mind-wandering, self-referential thought, mentalizing and semantic processing (Andrews-Hanna, 2012; Andrews-Hanna et al., 2014; Christoff et al., 2009).

Component 4 was comprised of deactivations in regions conventionally associated with the DMN, including the PCC, the AnG, and the amPFC. Such a deactivation departs from the proposed sub-network scheme discussed above (i.e. sensory, affective, cognitive, and motor sub-networks of pain; Davis & Moayedi, 2013; Wilcox et al., 2015). However, the DMN has also been implicated in pain and so its detection here is not entirely unexpected. In chronic pain disorders, for example, the DMN shows a number of anatomical-functional alterations, including fragmentation between frontal and posterior regions (Baliki et al., 2014), and strengthening of functional connections to aIns (Baliki et al., 2014; Loggia et al., 2013). In healthy individuals, heat-induced deactivations in several DMN regions have been reported (Kong et al., 2010), while some regions, like the hippocampus and precuneus, also predict pain ratings (in addition to stimulus intensity) by the magnitude of their deactivation (Atlas et al., 2014).

As others have argued, pain-induced deactivations in the DMN may be part of an attentional response to pain (Kucyi et al., 2013; Kucyi & Davis, 2015), where the DMN suppresses as attentional networks (e.g. Component 2) engage. This type of antagonistic relationship between the DMN and attentional networks has been documented extensively outside of pain imaging, along with the DMN’s “task-negative” tendencies (Anticevic et al., 2012; Peng et al., 2018). Future research would benefit from an analysis of DMN response to pain in the context of attentional manipulations. Alternatively, attention levels during a stimulus could be monitored to allow for an analysis of the relationships between DMN deactivation, DMN-Component 2 antagonism, pain perception and attention.

Also of note, several DMN regions, including the mPFC, hippocampus, and precuneus, have been associated with the regulation of pain (Goffaux et al., 2014; Schweinhardt & Bushnell, 2010). Their involvement implies a potential role of the DMN, which might accomplish regulation by interacting with the periaqueductal gray (PAG)—part of a descending pathway for pain control—through the mPFC (Kucyi et al., 2013). Thus, chronic pain disorders may be related, in part, to deficits in pain regulation caused by alterations to the DMN. This possibility requires further investigation and presents an important research objective due to its implications for chronic pain treatment.

Among intended applications of neuroimaging in the field of pain is the development of models to accurately classify an individual in pain. Previous attempts of this nature have adopted multivariate pattern analysis (MVPA; Haynes, 2015; van der Miesen et al., 2019). In brief, MVPA uses machine learning algorithms to model behavioural responses (either ordinal or continuous variables) as a function of multiple voxels (or “features”) considered simultaneously (Moayedi et al., 2018; van der Miesen et al., 2019); predictions or classifications of mental states are then generated on independent “testing” data based on model parameters learned in the “training” set (Rosa & Seymour, 2014). In one notable study applying MVPA, a network of regression weights distributed over pain regions (the “neurologic pain signature” or NPS) tracked physical pain intensity between individuals (Wager et al., 2013; Woo et al., 2015). Perhaps even more remarkable is that physical pain could be accurately distinguished from other types of pain (e.g., social; Wager et al., 2013).

In this study, regression models provided some insight into the capacity of networks detected through fMRI-CPCA to be used for pain prediction, as components 1, 2 and 4 were significantly associated with pain ratings both within and between participants. Importantly, this was not a predictive model (networks were used to model in-sample ratings with no predictions generated on new or held-out data), and the findings should not be interpreted as direct evidence of prediction ability. However, networks did show potential to be used in predictive analyses given that in-sample estimation was moderately accurate, and, importantly, results were achieved without any a priori selection of brain regions, reflecting a distinct advantage of fMRI-CPCA compared to other approaches.

Of all networks, Component 2 was most strongly related to pain perception; the relationship was positive and consistently accounted for the largest proportion of within- and between-subject variability in pain. The value of Component 2 for predicting pain is intuitive, insofar as brain regions included in this network represent sensory, affective, and cognitive dimensions of pain (Melzack & Casey, 1968). The DMN was also important for pain estimation, with the magnitude of its deactivations being significantly related to perceived pain intensity, both within and between participants. The relationships of both networks to pain are corroborated by previous work that has identified several Component 2 regions—including SII, aIns, dACC, left cerebellum, and IPL—and DMN regions—including hippocampus and precuneus—as explicit mediators of pain (i.e. they mediate the relationship between stimulus intensity and pain rating; Atlas et al., 2014).

The intensity of activation in Component 1 was unrelated to the intensity of perceived pain, mirroring the behaviour of SI itself, which codes pain information primarily in terms of sensory-discriminative attributes (Moulton et al., 2012). This aspect of Component 1 (i.e. its independence from pain perception) is corroborated by mediation analyses that demonstrate a preference of sensory cortex and cerebellum to stimulus intensity over pain report (Atlas et al., 2014), and implies that motor systems are mobilized in accordance with stimulus properties only; the perception of pain occurs elsewhere, and the intensity of motor commands is, on its own, an unreliable proxy for the intensity of that perception.

Despite significant associations, when converted into a classifier the model discriminated between pain and warmth with an accuracy of only 68.83%. While significantly greater than chance, sensitivity and specificity were low (estimated at 59.17% and 79.10%, respectively). Still, comparisons between components 1, 2 and 4 and the existing “neurological pain signature” (NPS) reveal a high degree of overlap. Common regions include aIns, pIns, supramarginal gyrus, thalamus, and IPL. Further, the NPS included negative predictive weights in regions that were deactivated in Component 4, including PCC, precuneus and mPFC (Wager et al., 2013). These anatomical similarities raise the possibility that accurate predictions of pain could be generated from components 1, 2 and 4 if specific regional activations (compared to an overall estimate of activation in the entire network) were accounted for using MVPA (Allefeld & Haynes, 2015). By avoiding spatial averaging, MVPA accounts for signal non-uniformities between voxels, and exploits these differences in response signal as a source of predictive information (Hebart & Baker, 2018).

Crucially, the predictive potential shown by components indicates that fMRI-CPCA may provide a useful tool for determining appropriate anatomical targets for MVPA. This is important because a critical step in the MVPA framework is the selection of “features” with which to train the machine learning algorithm (Rosa & Seymour, 2014). Features are typically a subset of voxels, whose activations will be related to the behavioural response by the algorithm (Allefeld & Haynes, 2015), and are selected from a region- or regions-of-interest (based on prior knowledge) or from the entire brain using dimensionality reduction techniques like PCA (van der Miesen et al., 2019). Restricting the analysis to relevant regions is important to mitigate the problem of features exceeding the number of observations, which may lead to overfitted models and interpretive challenges (van der Miesen et al., 2019). In the case of the NPS, features were selected a priori from a collection of well-established pain regions (Wager et al., 2013). By contrast, fMRI-CPCA would allow features to be selected from the predominant functional networks involved in pain perception, without relying on prior assumptions about relevant spatial or temporal response patterns. fMRI-CPCA thus provides an opportunity to select connectivity-based features (Rosa & Seymour, 2014) that are unbiased, data-driven and task-related.

As a final point, results from multiple regressions are not only relevant to pain prediction, but also reflect on network functions proposed earlier, specifically the roles of Component 2 and the DMN as attention networks. In the regressions, Component 2 and the DMN displayed opposite relationships to pain; higher pain was associated with greater activation in Component 2 but greater suppression in the DMN, both within and between participants. This is an extension of the pattern shown by estimated HDRs, where Component 2 became active during stimulation while the DMN became suppressed. Together, these findings suggest that Component 2 and the DMN assume an antagonistic configuration during pain, and that greater antagonism (i.e. greater separation in terms of activation) equates to a heightened perception of pain.

Based on neuroimaging literature, this antagonism is likely indicative of an ongoing attentional response. Component 2 included known salience network hubs in ACC and aIns, as well as cognitive pain regions associated with attention to pain, and the DMN’s role in attention has been well-documented. For example, the DMN tends to form anticorrelated relationships with frontoparietal attention networks during cognitively demanding tasks (Dixon et al., 2017; Menon, 2015; Sridharan et al., 2008), with greater deactivation predicting improved task performance (Anticevic et al., 2012). Furthermore, attention deficits are generally associated with increased DMN activation (Bonnelle et al., 2011; Weissman et al., 2006; Danckert & Merrifield, 2018). In the context of pain, DMN deactivation is especially pronounced when participants report attending to pain, and less so when participants mind-wander away from pain (Kucyi et al., 2013). Thus, the deactivation of DMN observed here likely signifies attention to pain. The simultaneous activation of Component 2—which included several regions known to be involved in attention—mirrors the stereo-typical antagonism between DMN and frontoparietal networks that underlies attention (Anticevic et al., 2012). In sum, these networks appear to contribute to pain perception by working together, in an anticorrelated fashion, as part of an attentional response process; the greater the attention, the greater the antagonism between networks and the greater the pain intensity.

It should be noted that the HDRs estimated by fMRI-CPCA are well-suited to making inferences about cognitive function; this is because fMRI-CPCA uses Finite Impulse Response (FIR)-basis sets to encode brain activity associated with task-timing, which are essentially dummy regressors for stimulus presentation timing that make no assumptions about the shape of the expected response. For this reason, the technique detects responses (and by extension, functional networks) elicited by cognitive processes that may go unnoticed in more traditional analyses, where the expected response is produced by convolving stimulus functions with canonical hemodynamic response functions (Henson & Friston, 2007; Henson et al., 2001; Lindquist, 2008; Lindquist et al., 2009). Detailed analysis of HDR shapes evoked in components 1, 2 and 4, under different experimental conditions, is therefore warranted to achieve a robust determination of network function.

Our study has a number of limitations to consider. First, we did not include a protocol for model validation when evaluating pain predictions and classifications made with multiple linear regression; the ability of the model to predict or classify pain in independent samples therefore remains unverified. Validation techniques—including cross-validation, hold-out validation, or bootstrapping—are common practice in decoding analyses to ensure that models generalize to out-of-sample data (Kohavi, 1995; van der Miesen et al., 2019). We did not apply these here because of properties of the data (primarily its small sample size of 30), which made a conventional approach challenging (e.g. some subjects never reported a stimulus as painful). Future research is needed to formally validate the pain predictive value of these networks.

In theory, the regression model obtained here is generalizable to new and independent data. The basic procedure would involve application of the regression coefficients obtained to the network activations estimated in a new individual to generate a predicted pain rating. This would require first obtaining an individual’s activation data during a thermal pain task, analysing their brain activity using fMRI-CPCA, and “classifying” the networks elucidated through in-house programs recently developed to determine which of the new individual’s networks most closely match with the networks that inform the current model (Percival et al., 2020). The HDR shapes associated with the correct networks would have to be averaged across an appropriate time interval (or an equivalent time interval to the current study) to generate estimates of network activation intensity, and the regression model obtained here would then be applied to network activation intensities to generate a predicted pain rating. The classification procedure referred to above has been utilized previously and involves correlating the loadings of networks obtained with the loadings of “template” images of networks, across a set of characteristic slices that define the individuality of a network. Each network identified in a new individual would have to be correlated with the template images of networks obtained in this study to determine the strongest matches. Ultimately, this classification procedure would aid in selecting networks whose activations (averaged over post-stimulus time) would then be subjected to the regression model in order to generate a pain prediction.

To address this first limitation (lack of model validation), future research may use larger datasets to re-conduct the current study with the addition of a validation protocol, or test the current model in new and independent data according to the procedure outlined above. That said, predictions based on these networks are likely to be improved if signal differences within sub-networks and regions of components are accounted for by using pattern-based analyses like MVPA, instead of constructing models based on a whole-network index of activation (i.e. estimated HDRs for entire networks).

Overall, this study has contributed to neuroimaging research on pain by elucidating three functional networks evoked by thermal stimulation: a sensorimotor response network for immediate pain avoidance (Component 1), a frontoparietal attention network mobilized by salience detection processes (Component 2), and the default-mode network (Component 4). Of these, attention and default-mode networks were related to pain perception both within and between participants. From a purely technical perspective, this study validates fMRI-CPCA within the domain of pain research for the first time, highlighting advantages compared to existing approaches, including that the parcellation of multiple task-related networks is accomplished without a priori selection of regions-of-interest (i.e. no assumptions about spatial properties of networks). Moreover, fMRI-CPCA does not rely on models that assume specific HDR shapes to identify task-related activity; instead, HDRs are predicted using FIR basis functions, which simply specify an interval during which task-relevant activity is expected to occur. In this way, fMRI-CPCA detects HDRs (and potentially networks) elicited by cognitive processes that may be unaccounted for in conventional analyses.

More generally, the findings obtained provide a foundation from which to further investigate these networks, their proposed functions and their pain predictive value. The networks identified (especially the attention and default-mode networks) may have implications for pain treatments, if they can be targeted successfully with strategies based on neuromodulation (Alo & Holsheimer, 2002), behavioural therapy (Eccleston et al., 2013), or real-time fMRI feedback (Chapin et al., 2012), for example. Further, validated pain predictions can be generated from these networks and potentially refined by applying MVPA within network boundaries. Patterns delineated through MVPA may ultimately serve as objective measures of pain, which are of crucial importance to effective pain management in patients unable to self-report their pain.