Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Multiple Functional Brain Networks Related to Pain Perception Revealed by fMRI

Multiple Functional Brain Networks Related to Pain Perception Revealed by fMRI The rise of functional magnetic resonance imaging (fMRI) has led to a deeper understanding of cortical processing of pain. Central to these advances has been the identification and analysis of “functional networks”, often derived from groups of pre-selected pain regions. In this study our main objective was to identify functional brain networks related to pain perception by examining whole-brain activation, avoiding the need for a priori selection of regions. We applied a data-driven technique—Constrained Principal Component Analysis for fMRI (fMRI-CPCA)—that identifies networks without assuming their anatomical or temporal properties. Open-source fMRI data collected during a thermal pain task (33 healthy participants) were subjected to fMRI-CPCA for network extraction, and networks were associated with pain perception by modelling subjective pain ratings as a function of network activation intensities. Three functional networks emerged: a sensorimotor response network, a salience-mediated attention network, and the default-mode network. Together, these networks constituted a brain state that explained variability in pain perception, both within and between individuals, demonstrating the potential of data-driven, whole-brain functional network techniques for the analysis of pain imaging data. . . . . . Keywords Functional MRI Functional brain networks Functional connectivity Pain Multivariate least-squares regression . . Principal component analysis Hemodynamic responses Attention Introduction The application of non-invasive neuroimaging techniques has greatly enhanced our neurobiological understanding of pain (Davis, 2011;May, 2008; Moayedi et al., 2018). Functional * John K. Kramer kramer@icord.org magnetic resonance imaging (fMRI) has played a particularly valuable role, leading to the discovery of a core set of Department of Psychology, University of British Columbia, 2136 regions—including the thalamus, the anterior cingulate, so- West Mall, Vancouver, BC V6T 1Z4, Canada matosensory, and insular cortices—that are consistently acti- BC Mental Health & Addictions Research Institute, BC Children’s vated by experimental pain (Davis & Moayedi, 2013;Iannetti Hospital Research Institute, 938 West 28th Ave, &Mouraux, 2010;Mouraux &Iannetti, 2018; Wilcox et al., Vancouver, BC V5Z 4H4, Canada 2015). ICORD, Blusson Spinal Cord Centre, 818 West 10th Ave, Traditionally, fMRI research on pain has relied extensively Vancouver, BC V5Z 1M9, Canada on mass-univariate analysis techniques to investigate the func- Department of Psychiatry, University of British Columbia, 2255 tional role of individual regions in generating the pain expe- Wesbrook Mall, Vancouver, BC V6T 2A1, Canada rience. More recently, functional connectivity (FC) tech- Department of Physical Therapy, University of British Columbia, niques, which examine temporal correlations between re- 2177 Wesbrook Mall, Vancouver, BC V6T 1Z3, Canada gions, have allowed researchers to determine how traditional Centre for Hip Health and Mobility, Robert H. N. Ho Research pain regions organize into larger networks. Characterizing Centre, 2635 Laurel St, Vancouver, BC V5Z 1M9, Canada such networks (in terms of both spatial organization and func- School of Kinesiology, University of British Columbia, 6081 tion) is an important objective because (1) largely distributed University Blvd, Vancouver, BC V6T 1Z1, Canada 156 Neuroinform (2022) 20:155–172 patterns of activation likely provide a more reliable “signa- In this paper, we used fMRI-CPCA to conduct a whole- ture” of pain than any local activation, where signatures have brain, data-driven extraction of functional networks involved the potential to be used in diagnosis and/or evaluations of treat- in pain. We analyzed a publicly available and previously pub- ment efficacy (van der Miesen et al., 2019), and (2) understand- lished dataset, posted on openneuro.org (accession number ing network functionality informs our basic understanding of ds000140; Gorgolewski, 2018; Woo et al., 2015), featuring existing treatments, for example, cognitive-behavioural therapies a thermal stimulation task. fMRI-CPCA delineated multiple, (Eccleston et al., 2013), as well as burgeoning treatment avenues dominant functional brain networks evoked by thermal stim- like neuromodulation (Alo & Holsheimer, 2002) and real-time ulation, obtaining estimates of their spatial configurations and fMRI feedback (Chapin et al., 2012). temporal response patterns. We then modelled subjective pain In pain research, FC techniques have shown traditional ratings as a linear function of multiple network activations, to pain regions to be organized into distinct functional networks verify the relevance of the networks detected to pain percep- serving sensory, emotional, cognitive or motor aspects of pain tion. Our fundamental goal was to identify the functional net- (Wilcox et al., 2015). However, FC studies have often relied works involved in processing noxious heat stimuli, explore on seed-based techniques, meaning that correlations between their responses and anatomy, and quantify their relationships brain regions are interrogated by selecting a voxel or region (a with pain perception. Based on research that has demonstrated “seed”) and modeling activity in other voxels as a function of the organization of pain regions into distinct networks at rest signal changes within the seed (Diano et al., 2016;Moayedi (described as sensory-discriminative, cognitive-evaluative, af- et al., 2018;Wilcox et al., 2015). Estimated model parameters fective-motivational, and motor networks; Davis & Moayedi, represent the strength of each voxel’s functional connection to 2013;Wilcox et al., 2015), we hypothesized similar network the seed and can be used to construct a map of intercorrelated configurations to be evoked during experimental thermal pain regions, that is to say, a functional network (Moayedi et al., based on our fMRI-CPCA analysis. 2018). Although powerful, this framework is limited by the regions (or seeds) inputted as regressors. It is therefore impor- tant to consider alternative methods that are data-driven, let- Materials and Methods ting functional networks emerge without relying on spatial (i.e. regions-of-interest) or temporal assumptions (i.e. pre- The original study by Woo et al. (2015) provides detailed supposing the shape of the response elicited, as is typically information on participants, study design and data collection. done in the univariate framework; Henson & Friston, 2007). Here, we provide only a brief description for clarity. One such alternative is Constrained Principal Component Analysis for fMRI (fMRI-CPCA). fMRI- Participants CPCA extracts functional brain networks from whole- brain Blood Oxygen Level Dependent (BOLD) signal data 33 healthy, right-handed adults (22 females, 11 males) participat- with variance constrained to that predictable from task ed in the study, with a mean age of 27.9 years (SD = 9.0 years). timing, and generates spatial maps, as well as estimates of All participants provided informed consent and reported no prior hemodynamic responses (HDRs) for each combination of history of psychiatric, neurological or pain disorders. Ethical re- subject, task condition and brain network. The technique is view and approval were provided by the Columbia University valuable in that it combines: (1) networks based on multi- Institutional Review Board (Protocol number AAAE3743). variate analyses, which interrogate the intercorrelated struc- Since the data were anonymized and we performed a secondary ture of task-based voxel data without submitting each voxel analysis, no local ethics review was required. to a separate statistical test as in univariate approaches (e.g. In our study, two participants (subjects 11 and 30) were seed-based connectivity techniques, where each voxel is excluded because they received too few trials under each ex- correlated to the seed), (2) networks extracted from BOLD perimental condition (defined below), creating problems for signal constrained to task-timing-related variance, which is the fMRI-CPCA algorithm. This left 31 participants to be useful because task-optimized networks can be more readily analysed. associated with cognitive and behavioural functions by analysing how network HDRs differ between task condi- Thermal Stimulation tions, and (3) data-driven network extraction, meaning that no assumptions about the spatial or temporal properties of To elicit pain, a thermode device was placed on the volar networks are formally defined. Spatial and temporal as- surface of the left forearm (TSA-II Neurosensory Analyzer sumptions are avoided by analyzing all voxels in the brain with a 16-mm Peltier thermode endplate, Medoc Advanced instead of selecting regions-of-interest and using a Finite Medical Systems). Thermal stimuli were delivered at specific Impulse Response (FIR) model of task-evoked HDRs in- temperatures for 12.5 s each, with 3 s of ramp-up, 7.5 s at the stead of assuming a particular HDR shape, respectively. target temperature, and 2 s of ramp-down. Temperature levels Neuroinform (2022) 20:155–172 157 ranged from 40.8 °C to 47.3 °C (study documentation and and the AC-PC plane was oriented horizontally. Slice-time participant results are available at https://openneuro.org/ correction was performed to mitigate the temporal lag in slice datasets/ds000140/versions/00001). acquisition across the 2-s TR, using slice 21 as a reference. Realignment algorithms were applied to counteract displace- fMRI Task ment of voxels due to head movement, and runs that exceeded movement parameter thresholds of 4.5 mm in either z, x, y Participants completed 9 separate functional scanning ses- direction, as well as pitch, yaw or roll, by at least 50 scans, sions. There were 3 types of sessions: “standard” runs, where were removed from the analysis (subject 10, run 5 and 6; pain stimulation was received passively; “regulate-up” runs, subject 2, run 5; subject 4, run 1, 4 and 5). For each partici- where participants were instructed to increase the intensity of pant, functional scans were co-registered to their correspond- pain by cognitive control; and their counterpart, “regulate- ing structural images, and structural T1 scans were segmented down” runs. The regulation manipulation was intended to en- into gray matter, white matter, cerebrospinal fluid, meninges gage supplementary brain systems for pain regulation. For the and skull components. Finally, raw functional data were nor- explicit purposes of our study, we focused on standard runs malized to MNI template space (with a voxel size of 3 × 3 × only. 3 mm) and smoothed with a 6 × 6 × 6 FWHM Gaussian Each standard run began with an 18-s fixation cross pre- kernel. sented on screen, followed by 11 consecutive trials. Each trial For detailed explanations of preprocessing methods, along was 33–41 s long and featured the same progression: 12.5 s of with specific versions of software tools used, refer to supple- thermal stimulation, 4.5–8.5 s (jittered) of pre-rate rest, 11 s of mentary materials. pain rating (completed on screen using a hand-held remote), and 5–9 s (jittered) of post-rate rest. The rating period in- Task-Based Whole-Brain Network Analysis volved two kinds of rating; first, participants decided whether a stimulus was painful or not (this phase lasted 4 s), then Constrained Principal Component Analysis (CPCA) is a sta- participants rated the intensity of their sensation on a Visual tistical technique that combines multivariate least-squares re- Analogue Scale from 0 to 200, where the interval 0–100 rep- gression with principal component analysis (Hunter & resented non-painful warmth, and 100–200 represented the Takane, 2002; Takane & Shibayama, 1991; Takane & intensity of a stimulus perceived as painful. The scale was Hunter, 2001). It can be used to perform whole-brain analyses presented on screen, and participants were instructed as to of fMRI BOLD signal data. When applied to fMRI, it iden- the meaning of each interval prior to scanning. The specific tifies multiple functional networks involved in a task and es- order of temperatures administered throughout each run can timates fluctuations in BOLD signal for each network, over a be found in Woo et al. (2015). For a schematic illustration of specified interval of time. Further statistical tests can be used task design, see Fig. 1. to quantify the interactions between networks, correlational relationships between network activation and behavioural Image Acquisition measures, and the effect of experimental manipulations on the activation of each network. Whole-brain functional images were collected on a 3 T Philips Broadly speaking, fMRI-CPCA involves two steps. First, Achieva TX scanner at Columbia University’s Program for multivariate least-squares multiple regression is used to isolate Imaging in Cognitive Science (PICS). Structural images were variance in BOLD signal that is predictable from the timing of collected with high-resolution T1 spoiled gradient recall im- stimulus presentation, after which the variance is said to be ages (SPGR), which allow for anatomical localization and “constrained” to task timing. This first step is referred to as the warping to standard space. For functional EPI image collec- external analysis. Second, a principal component analysis tion, the following scanning parameters were set: TR = (PCA) is conducted on the constrained portion of the variance 2000 ms, TE = 20 ms, field of view = 224 mm, 64 × 64 ma- in BOLD signal, and the extracted components represent sys- trix, 3 × 3 × 3 mm voxels, 42 interleaved slices, parallel im- tems of functionally interconnected voxels (i.e. functional aging, SENSE factor 1.5. E-Prime software (PST Inc.) was brain networks) related to the task. This step is referred as used to control stimulus presentation and collect behavioural the internal analysis. Importantly, applying PCA after the re- data. gression ensures that the networks identified are based on task-related information only. This is a defining feature of Preprocessing fMRI-CPCA and distinguishes it from other applications of PCA (or ICA) used in fMRI. In fMRI-CPCA, the variance For our analysis, all preprocessing was completed in SPM 12. shared between principal components and task timing is max- Structural and functional scans were reoriented manually, imized, thus avoiding any contamination of the solution by such that the origin was placed on the anterior commissure, variability that is not predictable from task timing. Ultimately, 158 Neuroinform (2022) 20:155–172 Fig. 1 Schematic illustration of task design adapted from Woo et al. pre-rate anticipation period (4.5–8.5 s, jittered), a 4 s rating period to (2015). Every run was preceded by an 18 s fixation cross presented on judge if the stimulus was painful or not, a 7 s pain rating period using a screen. Every trial began with a 12.5 s thermal stimulus, followed by a VAS scale, and a post rate rest period of 5–9 s (jittered) fMRI-CPCA outputs brain activity maps that can be overlaid zero is in all other cells—thus, the G matrix simply defines the on a structural image (for example, in applications like time intervals during which we expect to see task-relevant MRIcron [https://www.nitrc.org/projects/mricron]), as well activations. The number of rows in the G matrix will equal as estimated hemodynamic response shapes (plotted over the number of rows in the Z matrix, but the number of columns post-stimulus time) for each combination of network, is equal to the number of post-stimulus time points (time bins) subject and task condition. The next few paragraphs for which the BOLD signal is to be predicted, multiplied by will elaborate on specific matrices and equations the total number of conditions and the total number of sub- required to implement the analyses. jects. The G matrix is also standardized for each individual In order to perform the external analysis, two matrices must run. We then regress the Z matrix onto the G matrix, first be prepared. The Z matrix (or activation matrix) contains the BOLD data for all runs, with each voxel represented as a Z ¼ GC þ E; single column, and each full-brain scan represented as a single −1 where C =(G ′ G) G ′ Z is a matrix of timepoint- and voxel- row. In the current study, 31 subjects went through nine runs specific regression weights that satisfy the least-squares crite- each, with 209 scans per run. Six runs were removed due to rion. When C is applied to G it provides a matrix of BOLD excessive head movement (see section “2.5. Preprocessing”), signal values predicted from task-timing, Z or GC. E leaving a total of 42,427 rows (full brain scans) and 79,522 scans voxels columns (voxels) in the Z matrix. The mean value for each represents the residual signal (i.e. signal that is not predictable voxel was centered to zero for each run separately, and the from task-timing), which is disregarded in the rest of the anal- variables standardized (such that the standard deviations were ysis. As an additional note, E can be further analysed exactly set to one for each run separately). The G matrix (or design like GC; such an analysis would produce the dominant net- matrix) contains a Finite Impulse Response (FIR) model of the works that are not predictable from task timing, which may be BOLD signal based on stimulus presentation timing; unlike those engaged during off-task periods, or task on processes more conventional models, the FIR model does not impose a that span the whole series of trials but are not specifically predetermined HDR shape on the dataset (which is commonly elicited by the onset of tasks. This type of analysis was not assumed to aid in determining task-relevant activations in carried out here and is beyond the scope of this paper. BOLD signal). Instead, a value of one is placed into cells of The next stage of fMRI-CPCA is the internal analysis, which typically involves application of a principal component G for which the BOLD signal is to be estimated, and a value of Neuroinform (2022) 20:155–172 159 analysis (PCA) to the constrained, task-related signal (GC). did not substantially improve the solution, only marginally in- This identifies correlated structure underlying the voxel data, creasing variance explained and failingtodetectnew andinfor- grouping correlated voxels into components that represent mative regions/networks, insteadfragmentingnetworksprevi- functional brain networks. Importantly, these components will ously identified in the four-component solution. be optimized to be task-related, because GC contains task- After the external and internal analyses are complete, a related variance only. PCA is achieved through singular value final step is applied to produce estimates of HDR shapes as- decomposition of GC: sociated with each network. This is achieved by relating com- ponent scores (in matrix U) back to stimulus presentation timing (coded in G), and computing P such that: UDV ¼ SVDðÞ GC U ¼ GP; 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 where P contains “predictor weights”—these are weights that vectors. In matrix U, columns represent components, and estimate the intensity of each component for the time bins rows represent scans. The values in matrix U are “component specified in G. When plotted over post-stimulus time, predic- scores” and provide an indication or “score” of how important tor weights reveal the unique HDR shapes elicited by each each component is for each scan. In matrix V, columns repre- subject and condition within each network, for the specified sent components, and rows represent voxels. Cells of V can be pffiffiffiffi interval of time. In this study, predictor weights were averaged rescaled by VD= N to obtain “component loadings”—corre- over subjects before plotting. Further, predictor weights were lation coefficients indicating the correlation of task-related averaged over post-stimulus time to compute overall intensity BOLD signal in each voxel with the respective component values for network activation; more detail on this is provided scores. Voxels that are highly correlated with a given compo- below. nent’s “component scores” form the brain regions that define the functional network represented by that component. Preparation of G Notably, rescaling right singular vectors in V allows them to be interpreted as correlations between voxels and networks, The goal of the current study was to determine how the brain while also providing a better approximation of the inputted configures itself when processing pain, and to use the brain matrix GC by incorporating the variance accounted for by networks detected to generate a model of subjective pain per- each network. To visualize a brain activity map for each net- ception. Accordingly, we formatted the G matrix such that work, columns in rescaled V were overlaid on a brain template separate HDR shapes would be produced for high and low in MRIcron and thresholded to display only the voxels with temperature conditions. The division was based on the median the most dominant loadings (e.g. top 10% absolute values). In temperature administered across all trials, including regulation the current study, we orthogonally rotated and rescaled the V runs (the median temperature was 44.3 °C, any stimulus that matrix prior to display, using a varimax solution with 500 was equal to or less than 44.3 was assigned to the “low” iterations (Abdi & Williams, 2010; Bryant & Yarnold, 1995; temperature condition; the rest were “high”). We examined Kaiser, 1958). brain activity during thermal heat portions of the experimental PCA identifies a large number of components, but a select task only, and only included standard runs in the analysis to few can be extracted (the components that account for the least avoid capturing brain systems for cognitive self-regulation amount of variability are considered noise, or brain activity that is over pain perception. The task-relevant time interval (encoded unlikely to be reliable). Various methods for component selec- in G) was defined as the 16 s immediately following thermal tion exist; in this study we used the elbow method. This method stimulus presentation; in this way, the entire duration of the relies on visual inspection of the scree plot of singular values stimulus and 3.5 s thereafter were accounted for. Because each (Cattell, 1966; Cattell & Vogelmann, 1977). When plotted, sin- full-brain fMRI scan was completed in two seconds, HDRs gular values (which are contained in D) produce a line that grad- were estimated for eight post-stimulus time bins. The G matrix ually approaches zero as components account for less and less therefore consisted of 496 columns (2 conditions × 31 subjects variance. In the elbow method, components are selected for ex- × 8 time bins), and 42,427 rows (equal to the number of rows traction by locating the first abrupt increase in variance—relative in the Z matrix). 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 vari- Preparation of Network and Rating Data for Multiple ance (Kodinariya & Makwana, 2013). In this study, 4 compo- Regressions nents were retained and varimax rotated. We attempted addition- al analyses with a greater number of components retained, to For each network, the predictor weights produced in the final ensure the validity of the chosen threshold. Additional networks step of fMRI-CPCA define the unique HDR shapes associated 160 Neuroinform (2022) 20:155–172 with each condition, over the specified 16-s time interval. To sample was maintained between 45% and 55% (non- model pain perception from brain activation, it was preferable inclusive). to compute a single value that would capture the intensity of All brain networks detected by fMRI-CPCA (components the response. In this case, due to exploration of the HDR 1, 2, and 4) were then used to model the pain rating variable, shapes obtained, network activation intensity was estimated Rating∼1 þ Component1 þ Component2 þ Component4; by averaging predictor weights (i.e. estimated BOLD signal) over the entire post-stimulus time interval. This yielded 248 for every sample drawn (305 samples in total). This effective- estimates in total, one for each temperature category for each ly treated the regressors as random rather than fixed effects of the four networks detected by fMRI-CPCA for every (Fox, 2015), and it provided empirical bootstrap distributions participant. for relevant statistics like regression coefficients, R and mod- Pain ratings were subjected to a similar procedure: ratings el significance as determined by F-test, from which estimates associated with each temperature condition were averaged of each metric or model parameter could be obtained. over trials to obtain participant-specific estimates of pain per- Confidence intervals (95%) were calculated non- ception during high- and low-temperature stimuli, yielding 62 parametrically using percentiles (Fox, 2015). estimates in total. Model fit was further evaluated by determining the accura- cy with which fitted values distinguished between pain and Multiple Regressions warmth. This was done by converting the ratings predicted by the model (a continuous variable) into a categorical outcome, Two separate multiple linear regression analyses were con- pain or non-pain, based on the 100-point pain threshold spec- ducted on pain rating and network activation data. ified by the VAS scale. For every model, the predicted binary outcomes were compared to the true state of affairs in order to generate estimates of accuracy (the proportion of total cases Modelling Within-Subject Pain that were correctly classified as either pain or warmth), sensi- tivity (the proportion of pain cases that were correctly classi- The first of these modelled changes in perceived pain as a fied as pain) and specificity (the proportion of non-pain cases function of changes in the intensity of network activations. that were correctly classified as warmth). This provided em- The fundamental goal here was to examine how a change in pirical bootstrap distributions and corresponding estimates rating corresponded with changes in activation intensities be- (with percentile intervals) of model accuracy, sensitivity and tween high and low temperatures. All brain networks detected specificity. All regression analyses were completed in by fMRI-CPCA (component 3 was excluded because it MATLAB R2019a (scripts available from https://github. reflected a movement artifact, see section "3. Results") were com/MatteoDamascelli/Multiple-Functional-Brain- inputted as predictors to explain changes in pain rating: Networks-Related-to-Pain-Perception-Revealed-by-fMRI.). ΔRating∼1 þ ΔComponent1 þ ΔComponent2 þ ΔComponent4: Results To evaluate model fit beyond R , fitted values were plotted against, and correlated with, the response variable using Summary of fMRI-CPCA Output Pearson’s r correlation coefficient. The scree plot of singular values indicated that four compo- Modelling Between-Subject Pain nents should be extracted. Components 1, 2, 3, and 4 accounted for 21.47, 7.26, 4.95, and 3.74% of task-related To investigate the relationship between perceived pain and variance in BOLD signal, respectively. Component images network activation intensity across subjects, we applied a and estimated HDR shapes are displayed in Figs. 2, 3, 4, 5, bootstrap-like regression procedure. Samples of size n =31 along with network activation intensities for each temperature were drawn from the dataset of condition- and participant- category (box plots). specific estimates of pain ratings and network activation in- Component 1 was primarily comprised of a) motor areas, tensities (see section “2.8. Preparation of Network and Rating including the primary motor cortex (M1), supplementary mo- Data for Multiple Regressions”). Each participant contributed tor area (SMA) and cerebellum, b) visual areas, including the one pain rating (and its corresponding network intensities), lateral occipital cortex (LO), and c) the primary somatosenso- selected at random, and every sample was a near-balanced ry cortex (S1). Component 2 featured a frontoparietal activity combination of ratings greater than 100 (i.e. painful) and rat- pattern that included activation peaks in the anterior cingulate ings less than 100 (i.e. warm). The prevalence of pain in the cortex (ACC), dorsolateral prefrontal cortex (dlPFC), anterior Neuroinform (2022) 20:155–172 161 Fig. 2 a-c Component 1. a Three-dimensional rendering of Component 1 distributions of Component 1 BOLD signal across participants, for both (based on the top 10% of component loadings) and estimated HDRs high and low temperature stimuli. BOLD signal was first averaged over associated with this network over the course of one thermal stimulation the entire post-stimulus time interval for each participant and each con- trial. The red bar placed over x-axis tick labels indicates the duration of a dition. The mean is given by ×. c Horizontal cross-sections of Component thermal stimulus (12.5 s). Estimated HDRs were obtained by averaging 1 (only the top 10% of component loadings are shown). Positive loadings the FIR-based predictor weights for each condition level and plotting in red, threshold = 0.17, max = 0.33. No negative loadings. Blue values them as a function of post-stimulus time. Error bars given by standard indicate the MNI coordinate of each slice in the z direction error. HDR = hemodynamic response. b Boxplots illustrating the Fig. 3 a-c Component 2. a Three-dimensional rendering of Component 2 distributions of Component 2 BOLD signal across participants, for both (based on the top 10% of component loadings) and estimated HDRs high and low temperature stimuli. In both cases, BOLD signal was first associated with this network over the course of one thermal stimulation averaged over the entire post-stimulus time interval for each participant trial. The red bar placed over x-axis tick labels indicates the duration of a and each condition. The mean is given by ×. c Horizontal cross-sections thermal stimulus (12.5 s). Estimated HDRs were obtained by averaging of Component 2 (only the top 10% of component loadings are shown). the FIR-based predictor weights for each condition level and plotting Positive loadings in red, threshold = 0.09, max = 0.21. No negative load- them as a function of post-stimulus time. Error bars given by standard ings. Blue values indicate the MNI coordinate of each slice in the z error. HDR = hemodynamic response. b Boxplots illustrating the direction 162 Neuroinform (2022) 20:155–172 Fig. 4 a-c Component 3. a Three-dimensional rendering of Component 3 HDR = hemodynamic response. b Boxplots illustrating the distributions (based on the top 10% of component loadings) and estimated HDRs of Component 3 BOLD signal across participants, for both high and low associated with this network over the course of one thermal stimulation temperature stimuli. In both cases, BOLD signal was first averaged over trial. The red bar placed over x-axis tick labels indicates the duration of a the entire post-stimulus time interval for each participant and each con- thermal stimulus (12.5 s). Estimated HDRs were obtained by averaging dition. The mean is given by ×. c Horizontal cross-sections of Component the FIR-based predictor weights for each condition level and plotting 3 (only the top 10% of component loadings are shown). Negative load- them as a function of post-stimulus time. Blue coloring indicates negative ings in blue, threshold = −0.08, max = −0.16. Positive loadings in red, loadings; graphs should be interpreted as displaying the intensity of de- threshold 0.08, max = 0.09. Blue values indicate the MNI coordinate of activation instead of activation. Error bars given by standard error. each slice in the z direction Fig. 5 a-c Component 4. a Three-dimensional rendering of Component 4 HDR = hemodynamic response. b Boxplots illustrating the distributions (based on the top 10% of component loadings) and estimated HDRs of Component 4 BOLD signal across participants, for both high and low associated with this network over the course of one thermal stimulation temperature stimuli. In both cases, BOLD signal was first averaged over trial. The red bar placed over x-axis tick labels indicates the duration of a the entire post-stimulus time interval for each participant and each con- thermal stimulus (12.5 s). Estimated HDRs were obtained by averaging dition. The mean is given by ×. c Horizontal cross-sections of Component the FIR-based predictor weights for each condition level and plotting 4 (only the top 10% of component loadings are shown). Negative load- them as a function of post-stimulus time. Blue coloring indicates negative ings in blue, threshold = −0.06, max = −0.14. Positive loadings in red, loadings; graphs should be interpreted as displaying the intensity of de- threshold 0.06, max = 0.07. Blue values indicate the MNI coordinate of activation instead of activation. Error bars given by standard error. each slice in the z direction Neuroinform (2022) 20:155–172 163 Table 1 Within-subject pain: estimated regression coefficients and and posterior insula (aIns and pIns), and thalamus. related statistics Component 3 was limited to the outer edge of the cortex (specifically the frontal and occipital poles) and the longitu- Predictor b (S.E.) β t p value dinal fissure and was mostly composed of negative loadings (Intercept) 40.78 (7.19) 5.67 0.000 (i.e. it became deactivated during stimulation). This particu- Component 1 -138.94 (79.41) −0.29 −1.75 0.092 lar configuration was biologically untenable and resembled Component 2 201.39 (54.54) 0.65 3.69 0.001 no established networks; it was most likely summarizing Component 4 112.07 (40.77) 0.48 2.75 0.011 head movement that was coordinated with the application of thermal pain. Component 4 was characterized by deactivations b = unstandardized regression coefficient, β = standardized regression co- in areas conventionally associated with the default mode net- efficient, S.E. = standard error, p-values are two-tailed work, including the posterior cingulate cortex (PCC), medial prefrontal cortex (mPFC), precuneuous and angular gyrus consisted primarily of negative loadings (i.e. its intensity (AnG). Detailed anatomical descriptions for all networks are values represented deactivation, not activation, intensity). found in supplementary information (supplementary tables 1–3). To further evaluate model fit, changes in pain rating were plotted against changes predicted by the model in Fig. 6.The two variables were significantly correlated (r = 0.630, Regressions p < .001). Importantly, this model shows that the networks identified by fMRI-CPCA, as a whole, capture variations in To relate these networks back to pain perception, we modelled pain perception at the within-subject level. 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 rat- Between-Subject Pain ings and network activation intensities for each temperature category (i.e. high and low) within each participant. Bootstrapped regression models of pain ratings, with all func- Descriptive statistics (means and standard errors) for these tional networks inputted as predictors, explained 28.6% of the data are found in supplementary table 4. variance in pain rating data on average (R = 0.286, CI = [0.079, 0.475]), or 20.7% when adjusted for the num- 95% ber of predictors (R = 0.207). F-tests for the variance Within-Subject Pain adjusted accounted for by each model revealed that 65.9% of the time, For within-subject pain, the model included temperature- models were significant at the .05 level with a median p value dependent changes in network activation intensity for all func- of .023 (see Fig. 7). tional networks identified by fMRI-CPCA (components 1, 2 The accuracy of predicted pain ratings was evaluated by and 4), and explained 39.7% of the variance in temperature- averaging RMSE across all re-sampled models (RMSE = dependent changes in pain rating (R =0.397; F(3, 27) = 5.9, p 37.24; Fig. 7). Estimated regression coefficients, their stan- = .003), or 32.9% when adjusted for the number of predictors dard errors and confidence intervals are given in Table 2; (R = 0.329). This indicates that changes in pain ratings components 2 and 4 were the only significant predictors of adjusted are predictable from changes in BOLD signal in the functional pain rating (Component 2: β = 0.42, CI = [0.12, 0.64]; 95% networks identified, according to: Component 4: β =0.26, CI = [0.09, 0.44]). Further, stan- 95% dardized coefficients showed that Component 2 made the change in rating∼−138:94ðÞ C1 BOLD signal intensity change most important contribution to the model, and its activation þ 201:39ðÞ C2 BOLD signal intensity changeþ intensity was positively associated with pain ratings. By con- trast, Component 4 deactivation intensity was positively asso- 112:07ðÞ C4 BOLD signal intensity changeþ 40:78: ciated with pain ratings. Figure 8 provides a schematic sum- The accuracy of predicted scores was evaluated by taking mary of these relationships between networks and pain per- the standard deviation of the residuals or the Root-mean- ception. Overall, this model shows that—as a whole—the square error (RMSE), which was 18.27. As shown in networks delineated by fMRI-CPCA are sensitive to be- Table 1, components 2 and 4 were the only statistically sig- tween-subject variability in pain perception. nificant contributors (Component 2: β = 0.654, p =.001; To evaluate the model’s ability to differentiate be- Component 4: β = 0.482, p = .011). Also of note, tween painful and non-painful states, we converted true Component 2 predicted increases in pain based on increases ratings and predicted ratings into binary categories (i.e. in its activation, whereas Component 4 predicted increases in pain or non-pain, as described in method) and calculat- pain based on increases in its deactivation, given that it ed accuracy, sensitivity and specificity of classification 164 Neuroinform (2022) 20:155–172 Fig. 6 a-d Assessing model fit for within-subject pain. Temperature- model. d True change in pain ratings plotted against change in ratings dependent change in pain ratings plotted against change in BOLD signal predicted by the model. The strength and significance of the relationship for Component 1 (a), Component 2 (b), and Component 4 (c); true is given by Pearson r = 0.630, p < .001 changes are shown alongside predictions made by the linear regression for every resampled model; Supplementary Figure 1 predicted ratings tend to better approximate ratings be- demonstrates this procedure for one bootstrap sample. low the pain threshold than ratings above it, such that Empirical bootstrap distributions for accuracy, sensitivi- when ratings are converted to the binary variable “pain” ty and specificity are shown in Fig. 9. Estimates (ob- or “non-pain”, ratings above the threshold are more of- tained by averaging) were 68.83%, 59.17% and 77.10%, ten miscategorized than ratings below it. respectively. Confidence intervals show that only accu- racy and specificity were significantly higher than chance level (50%). This indicates that the overall ac- Discussion curacy of the model is driven by its specificity, or its ability to correctly identify non-pain (more precisely, In this study, distinct functional connectivity networks for specificity is the fraction of pain ratings below 100 pain were revealed by fMRI-CPCA. The networks encompass correctly classified as “non-pain”). It appears that the a variety of brain regions consistently active in response to Fig. 7 Assessing model fit for between-subject pain. Histograms, kernel 2 2 three figures: R ¼ 0.286, CI = [0.079, 0.475]; R = 0.207, 95% adj: density estimates, and average bootstrap estimates for R-squared, adjust- CI =[−0.023, 0.417]; RMSE = 37.238, CI = [31.247, 43.404]. 95% 95% ed R-squared, Root-mean-square Error (RMSE) and model significance For model significance: 65.9% of models were significant at the .05 level, afforded by F-test (determines if a model fits significantly better than one p = .023 median based on a constant term only). Means and percentile intervals for the first Neuroinform (2022) 20:155–172 165 Table 2 Between-subject pain: Unstandardized coefficient Standardized coefficient (β) estimated regression coefficients and related statistics Predictor Average Bootstrap Percentile CI (95%) Average Bootstrap Percentile CI (95%) Estimate (S.E.) Estimate (S.E.) (Intercept) 73.67 (16.81) [36.14, 103.13] Component 1 −48.95 (78.65) [−228.15, 87.88] −0.07 (0.11) [−0.30, 0.13] Component 2 172.26 (47.66) [52.76, 255.13] 0.42 (0.12) [0.12, 0.64] Component 4 76.24 (27.04) [24.54, 133.80] 0.26 (0.09) [0.09, 0.44] S.E. = standard error, i.e. the standard deviation of the corresponding bootstrap distribution pain, including MI, SMA, cerebellum and SI (Component 1), functional networks. Although the specific parcellation observed the ACC, insular cortex, and thalamus (Component 2), and here is unique, it is largely congruent with current perspectives mPFC, hippocampus, para-hippocampus and precuneus on pain-related networks. In particular, evidence from PET and (Component 4) (Apkarian et al., 2005; Atlas et al., 2014; fMRI suggests that pain-activated regions are segregated into at Schweinhardt & Bushnell, 2010). Within participants, chang- least four distinct sub-networks: a sensory network for stimulus es in perceived intensity related to low and high temperatures localization and intensity coding (Davis & Moayedi, 2013; were associated with the magnitude of change in BOLD Hofbauer et al., 2001;Peyronetal., 1999), an affective network across networks. While falling short of accurate classification, for generating the aversive, unpleasant quality of a stimulus the magnitude of BOLD activation in functional networks was (Davis & Moayedi, 2013;Peyron et al., 2000;Wilcoxetal., significantly associated with pain intensity between partici- 2015), a cognitive network for attending to, anticipating and pants. Future development of fMRI-CPCA in the context of remembering the stimulus (Davis & Moayedi, 2013;Peyron pain is warranted to further explore the brain in pain. et al., 1999;Wilcoxetal., 2015), and a network of motor regions Over and above capturing the activation of known pain re- for pain avoidance (Davis & Moayedi, 2013;Wilcoxet al., gions, fMRI-CPCA integrated these brain regions into multiple 2015). Fig. 8 a-b Component contributions to pain. a Schematic depiction of each component’s contribution to pain ratings, according to bootstrap estimates of standardized regression coefficients (beta weights). Lines are proportional to the strength of their relation to pain. Component 4 is characterized by deactivation instead of activation; thus, its β value captures the strength of the relationship between Component 4 deactivation and pain perception. b Histograms, kernel density estimates, and average bootstrap estimates for unstandardized regression coefficients, with means and CI bounds. Component number indicated in the top left corner 166 Neuroinform (2022) 20:155–172 Fig. 9 Classification performance. Histograms, kernel density estimates, and average bootstrap estimates of model accuracy (Mean = 68.83, acc. CI = [54.84, 83.87]), 95% sensitivity (Mean = 59.17, sens. CI = [33.33, 80.63]) and 95% specificity (Mean = 77.10, spec. CI = [57.14, 93.75]) 95% Component 1 (Sensorimotor Response) observed because of screen-related cues that coincided with stimulus presentation. Component 1, with prominent activation peaks in MI, SMA, and cerebellum, is most aptly described as a sensory and Component 2 (Attentional Pain Network) motor network. In the context of thermal stimulation, sensa- tion and motor output may be related to an instinctive In agreement with previous studies, Component 2 incorporat- flexing or bracing, or a desire to move, in response to in- ed a large number of regions involved in pain, and combined tense stimuli (Davis & Moayedi, 2013; Davis et al., 2002). sensory, affective and cognitive sub-networks (Davis & Hemodynamic response shapes (HDRs) for Component 1 Moayedi, 2013;Wilcox et al., 2015). For example, the most showed that activation was in fact exclusive to higher inten- prominent activation peaks were found in SII and posterior sity stimuli (temperatures above 44.3 °C). Component 1 also insula (pIns; sensory-discriminative regions), dACC, aIns, included prominent activations in SI, which, as a key corti- and thalamus (affective-motivational regions), and dlPFC cal aspect of the lateral nociceptive system, is one of the first and IPL (cognitive-evaluative regions; Peyron et al., 1999; recipients of ascending pain signals through the Wilcox et al., 2015). Based on this, Component 2 could reflect spinothalamic tract (Davis & Moayedi, 2013;Fomberstein a unification of sensory, affective and cognitive processes et al., 2013; Vierck et al., 2013; Yam et al., 2018). SI’s (Melzack & Casey, 1968) into a coordinated pain response. inclusion in Component 1 suggests that, during thermal The blending of sub-networks is likely facilitated by their stimulation, motor processes are prioritized and closely co- inter-connectivity at rest, provided by common nodes in ACC ordinated with sensory-discriminative functions (e.g. deter- and aIns that serve as relay sites between sub-networks mination of stimulus location and intensity). In theory, such (Wilcox et al., 2015). Importantly, the ACC and aIns are en- close communication would be necessary for an effective gaged in non-specific “salience detection”, where stimuli are pain avoidance response when stimulus intensity reaches selected based on their behavioural relevance, and attentional noxious levels (Postorino et al., 2017). This plausible role systems are primed to enable an effective response (Legrain of Component 1 in generating pain-induced motor com- et al., 2011; Menon, 2015). Such a “salience network” re- mands remains to be further explored; follow-up studies ceives axonal projections from sensory areas like the pIns, would benefit from monitoring physical movements in con- which are thought to provide the aIns with incoming sensory junction with other variables, allowing for the precise rela- information (Menon & Uddin, 2010). The pattern of activa- tionships between Component 1 activation intensity, stimu- tion observed in Component 2 captures both attentional sys- lus intensity (e.g. temperature), motion, and pain perception tems (i.e. the cognitive sub-network of pain) and sensory- to be determined. discriminative elements like SII and pIns. Thus, Component A novel observation from fMRI-CPCA is the temporal 2 may represent a salience network-mediated response to overlap between visual areas and sensory-motor coupling, ev- salience—in this case thermal stimulation—or more precisely idenced in Component 1. Their detection is likely an idiosyn- a sequential activation of sub-networks, i.e. sensory systems cratic capture of fMRI-CPCA, which avoids using regions-of- activate the salience network, which then activates cognitive interest to spatially constrain the analysis. In fact, the anatomy systems for sustained attention. The directionality of sub- of Component 1 replicates previous applications of fMRI- network relations is a matter of speculation, but it presents CPCA in other domains—specifically, it resembles a network an interesting question for future investigation. Additionally, consistently associated with sensorimotor response processes, the putative attentional function of Component 2 may be fur- featuring activations in lateralized MI, SMA, SI, cerebellum, ther explored by analyzing its pain-induced response during and visual areas including the lateral occipital cortex (LO; experimental manipulations of attentional demand or stimulus Goghari et al., 2017; Larivière et al., 2017; Metzak et al., salience; a larger effect of attention on network response than 2011). In this case, sensorimotor-visual coupling was likely stimulus temperature would suggest an attentional role. Neuroinform (2022) 20:155–172 167 Component 4 (Default-Mode Network) alterations to the DMN. This possibility requires further inves- tigation and presents an important research objective due to its The tendency for brain areas to become deactivated during a implications for chronic pain treatment. 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, Estimating Pain Within and Between Participants research has emerged documenting the overall functional con- tributions of the default-mode network (DMN) to human be- Among intended applications of neuroimaging in the field of havior, including its relevance to mind-wandering, self- pain is the development of models to accurately classify an referential thought, mentalizing and semantic processing individual in pain. Previous attempts of this nature have (Andrews-Hanna, 2012; Andrews-Hanna et al., 2014; adopted multivariate pattern analysis (MVPA; Haynes, Christoff et al., 2009). 2015; van der Miesen et al., 2019). In brief, MVPA uses Component 4 was comprised of deactivations in regions machine learning algorithms to model behavioural responses conventionally associated with the DMN, including the (either ordinal or continuous variables) as a function of mul- PCC, the AnG, and the amPFC. Such a deactivation departs tiple voxels (or “features”) considered simultaneously from the proposed sub-network scheme discussed above (i.e. (Moayedi et al., 2018; van der Miesen et al., 2019); predic- sensory, affective, cognitive, and motor sub-networks of pain; tions or classifications of mental states are then generated on Davis & Moayedi, 2013; Wilcox et al., 2015). However, the independent “testing” data based on model parameters learned DMN has also been implicated in pain and so its detection in the “training” set (Rosa & Seymour, 2014). In one notable here is not entirely unexpected. In chronic pain disorders, for study applying MVPA, a networkofregressionweights example, the DMN shows a number of anatomical-functional distributed over pain regions (the “neurologic pain sig- alterations, including fragmentation between frontal and pos- nature” or NPS) tracked physical pain intensity between terior regions (Baliki et al., 2014), and strengthening of func- individuals (Wager et al., 2013; Woo et al., 2015). tional connections to aIns (Baliki et al., 2014;Loggiaet al., Perhaps even more remarkable is that physical pain 2013). In healthy individuals, heat-induced deactivations in could be accurately distinguished from other types of several DMN regions have been reported (Kong et al., pain (e.g., social; Wager et al., 2013). 2010), while some regions, like the hippocampus and In this study, regression models provided some insight into precuneus, also predict pain ratings (in addition to stimulus the capacity of networks detected through fMRI-CPCA to be intensity) by the magnitude of their deactivation (Atlas et al., used for pain prediction, as components 1, 2 and 4 were sig- 2014). nificantly associated with pain ratings both within and be- As others have argued, pain-induced deactivations in the tween participants. Importantly, this was not a predictive mod- DMN may be part of an attentional response to pain (Kucyi el (networks were used to model in-sample ratings with no et al., 2013; Kucyi & Davis, 2015), where the DMN sup- predictions generated on new or held-out data), and the find- presses as attentional networks (e.g. Component 2) engage. ings should not be interpreted as direct evidence of prediction This type of antagonistic relationship between the DMN and ability. However, networks did show potential to be used in attentional networks has been documented extensively outside predictive analyses given that in-sample estimation was mod- of pain imaging, along with the DMN’s “task-negative” ten- erately accurate, and, importantly, results were achieved with- dencies (Anticevic et al., 2012; Peng et al., 2018). Future out any a priori selection of brain regions, reflecting a distinct research would benefit from an analysis of DMN response advantage of fMRI-CPCA compared to other approaches. to pain in the context of attentional manipulations. Of all networks, Component 2 was most strongly related to Alternatively, attention levels during a stimulus could be mon- pain perception; the relationship was positive and consistently itored to allow for an analysis of the relationships between accounted for the largest proportion of within- and between- DMN deactivation, DMN-Component 2 antagonism, pain subject variability in pain. The value of Component 2 for perception and attention. predicting pain is intuitive, insofar as brain regions included Also of note, several DMN regions, including the mPFC, in this network represent sensory, affective, and cognitive di- hippocampus, and precuneus, have been associated with the mensions of pain (Melzack & Casey, 1968). The DMN was regulation of pain (Goffaux et al., 2014; Schweinhardt & also important for pain estimation, with the magnitude of its Bushnell, 2010). Their involvement implies a potential role deactivations being significantly related to perceived pain in- of the DMN, which might accomplish regulation by tensity, both within and between participants. The relation- interacting with the periaqueductal gray (PAG)—part of a ships of both networks to pain are corroborated by previous descending pathway for pain control—through the mPFC work that has identified several Component 2 regions— (Kucyi et al., 2013). Thus, chronic pain disorders may be including SII, aIns, dACC, left cerebellum, and IPL—and related, in part, to deficits in pain regulation caused by DMN regions—including hippocampus and precuneus—as 168 Neuroinform (2022) 20:155–172 explicit mediators of pain (i.e. they mediate the relationship involved in pain perception, without relying on prior assump- between stimulus intensity and pain rating; Atlas et al., 2014). tions about relevant spatial or temporal response patterns. The intensity of activation in Component 1 was unrelated fMRI-CPCA thus provides an opportunity to select to the intensity of perceived pain, mirroring the behaviour of connectivity-based features (Rosa & Seymour, 2014)that SI itself, which codes pain information primarily in terms of are unbiased, data-driven and task-related. sensory-discriminative attributes (Moulton et al., 2012). This As a final point, results from multiple regressions are not aspect of Component 1 (i.e. its independence from pain per- only relevant to pain prediction, but also reflect on network ception) is corroborated by mediation analyses that demon- functions proposed earlier, specifically the roles of strate a preference of sensory cortex and cerebellum to stimu- Component 2 and the DMN as attention networks. In the lus intensity over pain report (Atlas et al., 2014), and implies regressions, Component 2 and the DMN displayed opposite that motor systems are mobilized in accordance with stimulus relationships to pain; higher pain was associated with greater properties only; the perception of pain occurs elsewhere, and activation in Component 2 but greater suppression in the the intensity of motor commands is, on its own, an unreliable DMN, both within and between participants. This is an exten- proxy for the intensity of that perception. sion of the pattern shown by estimated HDRs, where Despite significant associations, when converted into a Component 2 became active during stimulation while the classifier the model discriminated between pain and warmth DMN became suppressed. Together, these findings suggest with an accuracy of only 68.83%. While significantly greater that Component 2 and the DMN assume an antagonistic con- than chance, sensitivity and specificity were low (estimated at figuration during pain, and that greater antagonism (i.e. great- 59.17% and 79.10%, respectively). Still, comparisons be- er separation in terms of activation) equates to a heightened tween components 1, 2 and 4 and the existing “neurological perception of pain. pain signature” (NPS) reveal a high degree of overlap. Based on neuroimaging literature, this antagonism is likely Common regions include aIns, pIns, supramarginal gyrus, indicative of an ongoing attentional response. Component 2 thalamus, and IPL. Further, the NPS included negative pre- included known salience network hubs in ACC and aIns, as dictive weights in regions that were deactivated in Component well as cognitive pain regions associated with attention to 4, including PCC, precuneus and mPFC (Wager et al., 2013). pain, and the DMN’s role in attention has been well-docu- These anatomical similarities raise the possibility that accurate mented. For example, the DMN tends to form anticorrelated predictions of pain could be generated from components 1, 2 relationships with frontoparietal attention networks during and 4 if specific regional activations (compared to an overall cognitively demanding tasks (Dixon et al., 2017; Menon, estimate of activation in the entire network) were accounted 2015; Sridharan et al., 2008), with greater deactivation for using MVPA (Allefeld & Haynes, 2015). By avoiding predicting improved task performance (Anticevic et al., spatial averaging, MVPA accounts for signal non- 2012). Furthermore, attention deficits are generally associated uniformities between voxels, and exploits these differences with increased DMN activation (Bonnelle et al., 2011; in response signal as a source of predictive information Weissman et al., 2006; Danckert & Merrifield, 2018). In the (Hebart & Baker, 2018). context of pain, DMN deactivation is especially pronounced Crucially, the predictive potential shown by components when participants report attending to pain, and less so when indicates that fMRI-CPCA may provide a useful tool for de- participants mind-wander away from pain (Kucyi et al., 2013). termining appropriate anatomical targets for MVPA. This is Thus, the deactivation of DMN observed here likely signifies important because a critical step in the MVPA framework is attention to pain. The simultaneous activation of Component the selection of “features” with which to train the machine 2—which included several regions known to be involved in learning algorithm (Rosa & Seymour, 2014). Features are typ- attention—mirrors the stereo-typical antagonism between ically a subset of voxels, whose activations will be related to DMN and frontoparietal networks that underlies attention the behavioural response by the algorithm (Allefeld & (Anticevic et al., 2012). In sum, these networks appear to Haynes, 2015), and are selected from a region- or regions- contribute to pain perception by working together, in an of-interest (based on prior knowledge) or from the entire brain anticorrelated fashion, as part of an attentional response pro- using dimensionality reduction techniques like PCA (van der cess; the greater the attention, the greater the antagonism be- Miesen et al., 2019). Restricting the analysis to relevant re- tween networks and the greater the pain intensity. gions is important to mitigate the problem of features exceed- ing the number of observations, which may lead to overfitted Technical Considerations models and interpretive challenges (van der Miesen et al., 2019). In the case of the NPS, features were selected a priori Based on the literature discussed in sections above, it is pos- from a collection of well-established pain regions (Wager sible to infer the functionality of each network. However, et al., 2013). By contrast, fMRI-CPCA would allow features these inferences are speculative and are not necessarily vali- to be selected from the predominant functional networks dated by any direct experimental evidence obtained here; Neuroinform (2022) 20:155–172 169 instead they rely on prior notions about the functional contri- In theory, the regression model obtained here is generaliz- butions of regions or networks detected. Importantly, the able to new and independent data. The basic procedure would fMRI-CPCA framework provides an opportunity to more ro- involve application of the regression coefficients obtained to bustly characterize network function during a task. This is the network activations estimated in a new individual to gen- done by comparing the HDRs estimated for each network erate a predicted pain rating. This would require first obtaining across task conditions to determine how different combina- an individual’s activation data during a thermal pain task, tions of independent variables impact network behaviour. analysing their brain activity using fMRI-CPCA, and “classi- Statistical comparisons can be made using repeated- fying” the networks elucidated through in-house programs measures ANOVAs, with within-subject factors given by time recently developed to determine which of the new individual’s and independent variables of interest (e.g. temperature level in networks most closely match with the networks that inform this study). By carefully manipulating experimental condi- the current model (Percival et al., 2020). The HDR shapes tions, cognitive processes can be dissociated from each other, associated with the correct networks would have to be aver- and by interpreting main and interaction effects of factors on aged across an appropriate time interval (or an equivalent time HDRs, networks can be related to specific aspects of cognition interval to the current study) to generate estimates of network operationalized by task conditions. Comparisons can also be activation intensity, and the regression model obtained here made between populations of interest by adding between- would then be applied to network activation intensities to subject factors that define group membership. In this way, generate a predicted pain rating. The classification procedure network alterations or deficits associated with diagnostic referred to above has been utilized previously and involves categories—such as chronic pain disorders—can be correlating the loadings of networks obtained with the load- investigated. ings of “template” images of networks, across a set of charac- It should be noted that the HDRs estimated by fMRI-CPCA teristic slices that define the individuality of a network. Each are well-suited to making inferences about cognitive function; network identified in a new individual would have to be cor- this is because fMRI-CPCA uses Finite Impulse Response related with the template images of networks obtained in this (FIR)-basis sets to encode brain activity associated with task- study to determine the strongest matches. Ultimately, this timing, which are essentially dummy regressors for stimulus pre- classification procedure would aid in selecting networks sentation timing that make no assumptions about the shape of the whose activations (averaged over post-stimulus time) would expected response. For this reason, the technique detects re- then be subjected to the regression model in order to generate sponses (and by extension, functional networks) elicited by cog- apain prediction. nitive processes that may go unnoticed in more traditional anal- To address this first limitation (lack of model validation), yses, where the expected response is produced by convolving future research may use larger datasets to re-conduct the cur- rent study with the addition of a validation protocol, or test the stimulus functions with canonical hemodynamic response func- tions (Henson & Friston, 2007;Hensonetal., 2001; Lindquist, current model in new and independent data according to the 2008; Lindquist et al., 2009). Detailed analysis of HDR shapes procedure outlined above. That said, predictions based on evoked in components 1, 2 and 4, under different experimental these networks are likely to be improved if signal differences conditions, is therefore warranted to achieve a robust determina- within sub-networks and regions of components are tion of network function. accounted for by using pattern-based analyses like MVPA, instead of constructing models based on a whole-network in- Limitations dex of activation (i.e. estimated HDRs for entire networks). A second limitation is that we included stimuli not rated as Our study has a number of limitations to consider. First, we painful (based on the 100-point pain threshold specified by the did not include a protocol for model validation when evaluat- VAS scale) in both network extraction via fMRI-CPCA and ing pain predictions and classifications made with multiple regression models of pain ratings. For network extraction, this linear regression; the ability of the model to predict or classify means that networks delineated were composed of voxels that pain in independent samples therefore remains unverified. remained functionally-connected across non-painful and pain- Validation techniques—including cross-validation, hold-out ful stimulation; in this way, any voxels that became incorpo- validation, or bootstrapping—are common practice in rated into the networks—or any new networks that were decoding analyses to ensure that models generalize to out- formed—during painful stimuli only were potentially missed of-sample data (Kohavi, 1995; van der Miesen et al., 2019). by the analysis. For regression models, it raises the possibility We did not apply these here because of properties of the data that networks were related to warmth more so than pain per- (primarily its small sample size of 30), which made a conven- ception. This would be the case if model-based predictions of tional approach challenging (e.g. some subjects never reported ratings below the pain threshold (i.e. warmth) were consistent- a stimulus as painful). Future research is needed to formally ly better than those of ratings above the pain threshold (i.e. pain). validate the pain predictive value of these networks. 170 Neuroinform (2022) 20:155–172 Supplementary Information The online version contains supplementary Conclusion material available at https://doi.org/10.1007/s12021-021-09527-6. Overall, this study has contributed to neuroimaging research Code Availability Analyses were completed with software and custom on pain by elucidating three functional networks evoked by code available and freely accessible online. Specific links are provided in the Information Sharing Statement. thermal stimulation: a sensorimotor response network for im- mediate pain avoidance (Component 1), a frontoparietal atten- Author Contribution Matteo Damascelli: data curation, formal analysis, tion network mobilized by salience detection processes funding acquisition, investigation, methodology, software, visualization, (Component 2), and the default-mode network (Component writing – original draft. Todd S. Woodward: conceptualization, investi- 4). Of these, attention and default-mode networks were related gation, methodology, resources, software, supervision, writing – review & editing. Nicole Sanford: software. Hafsa B. Zahid: software. Ryan Lim: to pain perception both within and between participants. From methodology, software. Alexander Scott: conceptualization, funding ac- a purely technical perspective, this study validates fMRI- quisition, investigation, writing – review & editing. John K. Kramer: CPCA within the domain of pain research for the first time, conceptualization, funding acquisition, investigation, methodology, re- highlighting advantages compared to existing approaches, in- sources, supervision, writing – review & editing. cluding that the parcellation of multiple task-related networks Funding This research was partially funded by the Natural Sciences and is accomplished without a priori selection of regions-of- Engineering Research Council (NSERC; grant numbers 5456 and interest (i.e. no assumptions about spatial properties of net- RGPAS-2017-507820). works). Moreover, fMRI-CPCA does not rely on models that assume specific HDR shapes to identify task-related activity; Data Availability Data are available on a public repository online, and are instead, HDRs are predicted using FIR basis functions, which freely accessible. A link can be found in the Information Sharing Statement. 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 Declarations may be unaccounted for in conventional analyses. Ethics Approval For the original data, ethical review and approval were More generally, the findings obtained provide a foundation provided by the Columbia University Institutional Review Board from which to further investigate these networks, their proposed (Protocol number AAAE3743). Since these data were anonymized and functions and their pain predictive value. The networks identified we performed a secondary analysis, no local ethics review was required. (especially the attention and default-mode networks) may have implications for pain treatments, if they can be targeted success- Consent to Participate All participants provided informed consent. fully with strategies based on neuromodulation (Alo & Competing Interests The authors declare no competing interests. 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 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adap- from these networks and potentially refined by applying tation, distribution and reproduction in any medium or format, as long as MVPA within network boundaries. Patterns delineated through you give appropriate credit to the original author(s) and the source, pro- MVPA may ultimately serve as objective measures of pain, vide a link to the Creative Commons licence, and indicate if changes were which are of crucial importance to effective pain management made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a in patients unable to self-report their pain. credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Information Sharing Statement The thermal stimulation fMRI data that support the findings of References this study were originally published by Woo et al. (2015)and are available from openneuro.org, https://openneuro.org/datasets/ Abdi, H., & Williams, L. J. (2010). Principal component analysis. Wiley ds000140/versions/00001. Accession number ds000140. Interdisciplinary Reviews: Computational Statistics, 2(4), 433–459. The fMRI-CPCA program is available from https://www. Allefeld, C., & Haynes, J. D. (2015). Multi-voxel pattern analysis. In A. nitrc.org/projects/fmricpca; codes are implemented in W. Toga (Ed.), Brain mapping: An encyclopedic reference (pp. 641–646). Academic Press: Elsevier. MATLAB using a graphical user interface. Custom code for Alo, K. M., & Holsheimer, J. (2002). New trends in neuromodulation for within- and between-subject regression models can be accessed the management of neuropathic pain. Neurosurgery, 50(4), 690– from GitHub, https://github.com/MatteoDamascelli/Multiple- Functional-Brain-Networks-Related-to-Pain-Perception- Andrews-Hanna, J. R. (2012). The brain’s default network and its adap- tive role in internal mentation. Neuroscientist, 18(3), 251–270. Revealed-by-fMRI. Neuroinform (2022) 20:155–172 171 Andrews-Hanna, J. R., Smallwood, J., & Spreng, R. N. (2014). The Goghari, V. M., Sanford, N., Spilka, M. J., & Woodward, T. S. (2017). Task-related functional connectivity analysis of emotion discrimina- default network and self-generated thought: Component processes, dynamic control, and clinical relevance. Annals of the New York tion in a family study of schizophrenia. Schizophrenia Bulletin, Academy of Sciences, 1316(1), 29–52. 43(6), 1348–1362. Anticevic, A., Cole, M. W., Murray, J. D., Corlett, P. R., Wang, X. J., & Gorgolewski, C. (2018). Distinct brain systems mediate the effects of Krystal, J. H. (2012). The role of default network deactivation in nociceptive input and self-regulation on pain (00001) [data set]. cognition and disease. Trends in Cognitive Sciences, 16(12), 584– Retrieved from https://openneuro.org/datasets/ds000140/versions/ 00001. AccessedJune2018 Apkarian, A. V., Bushnell, M. C., Treede, R. D., & Zubieta, J. K. (2005). Haynes, J. D. (2015). A primer on pattern-based approaches to fMRI: Human brain mechanisms of pain perception and regulation in Principles, pitfalls, and perspectives. Neuron, 87(2), 257–270. health and disease. European journal of pain, 9(4), 463–484. Hebart,M.N.,&Baker,C.I.(2018).Deconstructingmultivariate Atlas, L. Y., Lindquist, M. A., Bolger, N., & Wager, T. D. (2014). Brain decoding for the study of brain function. Neuroimage, 180,4–18. mediators of the effects of noxious heat on pain. PAIN®, 155(8), Henson, R., & Friston, K. (2007). Convolution models for fMRI. 1632–1648. Statistical parametric mapping: The analysis of functional brain Baliki, M. N., Mansour, A. R., Baria, A. T., & Apkarian, A. V. (2014). images,178–192. Functional reorganization of the default mode network across chron- Henson, R., Rugg, M. D., & Friston, K. J. (2001). The choice of basis ic pain conditions. PLoS ONE, 9(9), e106133. functions in event-related fMRI. NeuroImage, 13(6), 149–149. Bonnelle, V., Leech, R., Kinnunen, K. M., Ham, T. E., Beckmann, C. F., Hofbauer, R. K., Rainville, P., Duncan, G. H., & Bushnell, M. C. (2001). De Boissezon, X., et al. (2011). Default mode network connectivity Cortical representation of the sensory dimension of pain. Journal of predicts sustained attention deficits after traumatic brain injury. The Neurophysiology, 86(1), 402–411. Journal of Neuroscience, 31(38), 13442–13451. Hunter, M. A., & Takane, Y. (2002). Constrained principal component Bryant, F. B., & Yarnold, P. R. (1995). Principal-components analysis analysis: Various applications. Journal of Educational and and exploratory and confirmatory factor analysis. In L. G. Grimm & Behavioral Statistics, 27(2), 105–145. P. R. Yarnold (Eds.), Reading and understanding multivariate Iannetti, G. D., & Mouraux, A. (2010). From the neuromatrix to the pain statistics (p. 99–136). American Psychological Association. matrix (and back). Experimental Brain Research, 205(1), 1–12. Cattell, R. B. (1966). The scree test for the number of factors. Kaiser, H. F. (1958). The varimax criterion for analytic rotation in factor Multivariate Behavioral Research, 1(2), 245–276. analysis. Psychometrika, 23(3), 187–200. Cattell, R. B., & Vogelmann, S. (1977). A comprehensive trial of the scree and KG criteria for determining the number of factors. Kodinariya, T. M., & Makwana, P. R. (2013). Review on determining Multivariate Behavioral Research, 12(3), 289–325. number of cluster in K-means clustering. International Journal, Chapin, H., Bagarinao, E., & Mackey, S. (2012). Real-time fMRI applied 1(6), 90–95. to pain management. Neuroscience Letters, 520(2), 174–181. Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy Christoff, K., Gordon, A. M., Smallwood, J., Smith, R., & Schooler, J. W. estimation and model selection. International Joint Conference on (2009). Experience sampling during fMRI reveals default network Artificial Intelligence, 14(2), 1137–1145. and executive system contributions to mind wandering. Proceedings Kong, J., Loggia, M. L., Zyloney, C., Tu, P., LaViolette, P., & Gollub, R. of the National Academy of Sciences, 106(21), 8719–8724. L. (2010). Exploring the brain in pain: Activations, deactivations Danckert, J., & Merrifield, C. (2018). Boredom, sustained attention and and their relation. Pain, 148(2), 257–267. the default mode network. Experimental Brain Research, 236(9), Kucyi, A., & Davis, K. D. (2015). The dynamic pain connectome. Trends 2507–2518. in neurosciences, 38(2), 86–95. Davis, K. D. (2011). Neuroimaging of pain: What does it tell us? Current Kucyi, A., Salomons, T. V., & Davis, K. D. (2013). Mind wandering Opinion in Supportive and Palliative Care, 5(2), 116–121. away from pain dynamically engages antinociceptive and default Davis, K. D., & Moayedi, M. (2013). Central mechanisms of pain re- mode brain networks. Proceedings of the National Academy of vealed through functional and structural MRI. Journal of Sciences, 110(46), 18692–18697. Neuroimmune Pharmacology, 8(3), 518–534. Larivière, S., Lavigne, K. M., Woodward, T. S., Gerretsen, P., Graff- Davis, K. D., Pope, G. E., Crawley, A. P., & Mikulis, D. J. (2002). Neural Guerrero, A., & Menon, M. (2017). Altered functional connectivity correlates of prickle sensation: A percept-related fMRI study. in brain networks underlying self-referential processing in delusions Nature Neuroscience, 5(11), 1121–1122. of reference in schizophrenia. Psychiatry Research: Neuroimaging, Diano, M., D’Agata, F., Cauda, F., Costa, T., Geda, E., Sacco, K., Duca, 263,32–43. S., Torta, D. M., & Geminiani, G. C. (2016). Cerebellar clustering Legrain, V., Iannetti, G. D., Plaghki, L., & Mouraux, A. (2011). The pain and functional connectivity during pain processing. Cerebellum, matrix reloaded: A salience detection system for the body. Progress 15(3), 343–356. in Neurobiology, 93(1), 111–124. Dixon, M. L., Andrews-Hanna, J. R., Spreng, R. N., Irving, Z. C., Mills, Lindquist, M. A. (2008). The statistical analysis of fMRI data. Statistical C., Girn, M., & Christoff, K. (2017). Interactions between the de- Science, 23(4), 439–464. fault network and dorsal attention network vary across default sub- Lindquist, M. A., Loh, J. M., Atlas, L. Y., & Wager, T. D. (2009). systems, time, and cognitive states. Neuroimage, 147,632–649. Modeling the hemodynamic response function in fMRI: Eccleston, C., Morley, S. J., & Williams, A. D. C. (2013). Psychological Efficiency, bias and mis-modeling. Neuroimage, 45(1), S187–S198. approaches to chronic pain management: Evidence and challenges. Loggia, M. L., Kim, J., Gollub, R. L., Vangel, M. G., Kirsch, I., Kong, J., British Journal of Anaesthesia, 111(1), 59–63. Wasan, A. D., & Napadow, V. (2013). Default mode network con- Fomberstein, K., Qadri, S., & Ramani, R. (2013). Functional MRI and nectivity encodes clinical pain: An arterial spin labeling study. pain. Current Opinion in Anesthesiology, 26(5), 588–593. PAIN®, 154(1), 24–33. Fox, J. (2015). Applied regression analysis and generalized linear May, A. (2008). Chronic pain may change the structure of the brain. models. Sage Publications. PAIN®, 137(1), 7–15. Goffaux, P., Girard-Tremblay, L., Marchand, S., Daigle, K., & Whittingstall, K. (2014). Individual differences in pain sensitivity Melzack, R., & Casey, K. L. (1968). Sensory, motivational, and central control determinants of pain: A new conceptual model. The Skin vary as a function of precuneus reactivity. Brain Topography, 27(3), 366–374. Senses, 1,423–443. 172 Neuroinform (2022) 20:155–172 Menon, V. (2015). Salience network. In A. W. Toga (Ed.), Brain map- Schweinhardt, P., & Bushnell, M. C. (2010). Pain imaging in health and disease—How far have we come? The Journal of Clinical ping: An encyclopedic reference (pp. 597–611). Academic Press: Elsevier. Investigation, 120(11), 3788–3797. Shulman, G. L., Fiez, J. A., Corbetta, M., Buckner, R. L., Miezin, F. M., Menon, V., & Uddin, L. Q. (2010). Saliency, switching, attention and Raichle, M. E., & Petersen, S. E. (1997). Common blood flow control: A network model of insula function. Brain Structure & changes across visual tasks: II. Decreases in cerebral cortex. Function, 214(5–6), 655–667. Journal of Cognitive Neuroscience, 9(5), 648–663. Metzak, P. D., Riley, J. D., Wang, L., Whitman, J. C., Ngan, E. T., & Sridharan, D., Levitin, D. J., & Menon, V. (2008). A critical role for the Woodward, T. S. (2011). Decreased efficiency of task-positive and right fronto-insular cortex in switching between central-executive task-negative networks during working memory in schizophrenia. and default-mode networks. Proceedings of the National Academy Schizophrenia Bulletin, 38(4), 803–813. of Sciences, 105(34), 12569–12574. Moayedi, M., Salomons, T. V., & Atlas, L. Y. (2018). Pain neuroimaging Takane, Y., & Hunter, M. A. (2001). Constrained principal component in humans: A primer for beginners and non-imagers. The Journal of analysis: A comprehensive theory. AAECC, 12(5), 391–419. Pain, 19(9), 961–9e1. Takane, Y., & Shibayama, T. (1991). Principal component analysis with Moulton, E. A., Pendse, G., Becerra, L. R., & Borsook, D. (2012). BOLD external information on both subjects and variables. Psychometrika, responses in somatosensory cortices better reflect heat sensation 56(1), 97–120. than pain. The Journal of Neuroscience, 32(17), 6024–6031. van der Miesen, M. M., Lindquist, M. A., & Wager, T. D. (2019). Mouraux, A., & Iannetti, G. D. (2018). The search for pain biomarkers in Neuroimaging-based biomarkers for pain: State of the field and the human brain. Brain, 141(12), 3290–3307. current directions. Pain Reports, 4(4). Peng, K., Steele, S. C., Becerra, L., & Borsook, D. (2018). Brodmann Vierck, C.J.,Whitsel,B.L.,Favorov,O.V.,Brown,A.W.,& area 10: collating, integrating and high level processing of Tommerdahl, M. (2013). Role of primary somatosensory cortex in nociception and pain. Progress in neurobiology, 161,1–22. the coding of pain. PAIN®, 154(3), 334–344. Percival, C.M., Zahid, H.B., & Woodward, T. S. (2020). CNoS-Lab/ Wager, T. D., Atlas, L. Y., Lindquist, M. A., Roy, M., Woo, C. W., & Woodward_Atlas. Zenodo. https://doi.org/10.5281/zenodo. Kross, E. (2013). An fMRI-based neurologic signature of physical pain. The New England Journal of Medicine, 368(15), 1388–1397. Peyron, R., García-Larrea, L., Grégoire, M. C., Costes, N., Convers, P., Weissman, D. H., Roberts, K. C., Visscher, K. M., & Woldorff, M. G. Lavenne, F., Mauguière, F., Michel, D., & Laurent, B. (1999). (2006). The neural bases of momentary lapses in attention. Nature Haemodynamic brain responses to acute pain in humans: Sensory Neuroscience, 9(7), 971–978. and attentional networks. Brain, 122(9), 1765–1780. Wilcox, C. E., Mayer, A. R., Teshiba, T. M., Ling, J., Smith, B. W., Peyron, R., Laurent, B., & Garcia-Larrea, L. (2000). Functional imaging Wilcox, G. L., & Mullins, P. G. (2015). The subjective experience of brain responses to pain. A review and meta-analysis (2000). of pain: An FMRI study of percept-related models and functional Neurophysiologie Clinique/Clinical Neurophysiology, 30(5), 263– connectivity. Pain Medicine, 16(11), 2121–2133. Woo, C. W., Roy, M., Buhle, J. T., & Wager, T. D. (2015). Distinct brain Postorino, M., May, E. S., Nickel, M. M., Tiemann, L., & Ploner, M. systems mediate the effects of nociceptive input and self-regulation (2017). Influence of pain on motor preparation in the human brain. on pain. PLoS Biology, 13(1), e1002036. Journal of Neurophysiology, 118(4), 2267–2274. Yam, M. F., Loh, Y. C., Tan, C. S., Khadijah Adam, S., Abdul Manan, N., & Basir, R. (2018). General pathways of pain sensation and the Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, major neurotransmitters involved in pain regulation. International D. A., & Shulman, G. L. (2001). A default mode of brain function. Journal of Molecular Sciences, 19(8), 2164. Proceedings of the National Academy of Sciences, 98(2), 676–682. Rosa, M. J., & Seymour, B. (2014). Decoding the matrix: Benefits and limitations of applying machine learning algorithms to pain neuro- Publisher’sNote Springer Nature remains neutral with regard to jurisdic- imaging. Pain, 155(5), 864–867. tional claims in published maps and institutional affiliations. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Neuroinformatics Springer Journals

Multiple Functional Brain Networks Related to Pain Perception Revealed by fMRI

Loading next page...
 
/lp/springer-journals/multiple-functional-brain-networks-related-to-pain-perception-revealed-WysQxuqxu1

References (92)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2021
ISSN
1539-2791
eISSN
1559-0089
DOI
10.1007/s12021-021-09527-6
Publisher site
See Article on Publisher Site

Abstract

The rise of functional magnetic resonance imaging (fMRI) has led to a deeper understanding of cortical processing of pain. Central to these advances has been the identification and analysis of “functional networks”, often derived from groups of pre-selected pain regions. In this study our main objective was to identify functional brain networks related to pain perception by examining whole-brain activation, avoiding the need for a priori selection of regions. We applied a data-driven technique—Constrained Principal Component Analysis for fMRI (fMRI-CPCA)—that identifies networks without assuming their anatomical or temporal properties. Open-source fMRI data collected during a thermal pain task (33 healthy participants) were subjected to fMRI-CPCA for network extraction, and networks were associated with pain perception by modelling subjective pain ratings as a function of network activation intensities. Three functional networks emerged: a sensorimotor response network, a salience-mediated attention network, and the default-mode network. Together, these networks constituted a brain state that explained variability in pain perception, both within and between individuals, demonstrating the potential of data-driven, whole-brain functional network techniques for the analysis of pain imaging data. . . . . . Keywords Functional MRI Functional brain networks Functional connectivity Pain Multivariate least-squares regression . . Principal component analysis Hemodynamic responses Attention Introduction The application of non-invasive neuroimaging techniques has greatly enhanced our neurobiological understanding of pain (Davis, 2011;May, 2008; Moayedi et al., 2018). Functional * John K. Kramer kramer@icord.org magnetic resonance imaging (fMRI) has played a particularly valuable role, leading to the discovery of a core set of Department of Psychology, University of British Columbia, 2136 regions—including the thalamus, the anterior cingulate, so- West Mall, Vancouver, BC V6T 1Z4, Canada matosensory, and insular cortices—that are consistently acti- BC Mental Health & Addictions Research Institute, BC Children’s vated by experimental pain (Davis & Moayedi, 2013;Iannetti Hospital Research Institute, 938 West 28th Ave, &Mouraux, 2010;Mouraux &Iannetti, 2018; Wilcox et al., Vancouver, BC V5Z 4H4, Canada 2015). ICORD, Blusson Spinal Cord Centre, 818 West 10th Ave, Traditionally, fMRI research on pain has relied extensively Vancouver, BC V5Z 1M9, Canada on mass-univariate analysis techniques to investigate the func- Department of Psychiatry, University of British Columbia, 2255 tional role of individual regions in generating the pain expe- Wesbrook Mall, Vancouver, BC V6T 2A1, Canada rience. More recently, functional connectivity (FC) tech- Department of Physical Therapy, University of British Columbia, niques, which examine temporal correlations between re- 2177 Wesbrook Mall, Vancouver, BC V6T 1Z3, Canada gions, have allowed researchers to determine how traditional Centre for Hip Health and Mobility, Robert H. N. Ho Research pain regions organize into larger networks. Characterizing Centre, 2635 Laurel St, Vancouver, BC V5Z 1M9, Canada such networks (in terms of both spatial organization and func- School of Kinesiology, University of British Columbia, 6081 tion) is an important objective because (1) largely distributed University Blvd, Vancouver, BC V6T 1Z1, Canada 156 Neuroinform (2022) 20:155–172 patterns of activation likely provide a more reliable “signa- In this paper, we used fMRI-CPCA to conduct a whole- ture” of pain than any local activation, where signatures have brain, data-driven extraction of functional networks involved the potential to be used in diagnosis and/or evaluations of treat- in pain. We analyzed a publicly available and previously pub- ment efficacy (van der Miesen et al., 2019), and (2) understand- lished dataset, posted on openneuro.org (accession number ing network functionality informs our basic understanding of ds000140; Gorgolewski, 2018; Woo et al., 2015), featuring existing treatments, for example, cognitive-behavioural therapies a thermal stimulation task. fMRI-CPCA delineated multiple, (Eccleston et al., 2013), as well as burgeoning treatment avenues dominant functional brain networks evoked by thermal stim- like neuromodulation (Alo & Holsheimer, 2002) and real-time ulation, obtaining estimates of their spatial configurations and fMRI feedback (Chapin et al., 2012). temporal response patterns. We then modelled subjective pain In pain research, FC techniques have shown traditional ratings as a linear function of multiple network activations, to pain regions to be organized into distinct functional networks verify the relevance of the networks detected to pain percep- serving sensory, emotional, cognitive or motor aspects of pain tion. Our fundamental goal was to identify the functional net- (Wilcox et al., 2015). However, FC studies have often relied works involved in processing noxious heat stimuli, explore on seed-based techniques, meaning that correlations between their responses and anatomy, and quantify their relationships brain regions are interrogated by selecting a voxel or region (a with pain perception. Based on research that has demonstrated “seed”) and modeling activity in other voxels as a function of the organization of pain regions into distinct networks at rest signal changes within the seed (Diano et al., 2016;Moayedi (described as sensory-discriminative, cognitive-evaluative, af- et al., 2018;Wilcox et al., 2015). Estimated model parameters fective-motivational, and motor networks; Davis & Moayedi, represent the strength of each voxel’s functional connection to 2013;Wilcox et al., 2015), we hypothesized similar network the seed and can be used to construct a map of intercorrelated configurations to be evoked during experimental thermal pain regions, that is to say, a functional network (Moayedi et al., based on our fMRI-CPCA analysis. 2018). Although powerful, this framework is limited by the regions (or seeds) inputted as regressors. It is therefore impor- tant to consider alternative methods that are data-driven, let- Materials and Methods ting functional networks emerge without relying on spatial (i.e. regions-of-interest) or temporal assumptions (i.e. pre- The original study by Woo et al. (2015) provides detailed supposing the shape of the response elicited, as is typically information on participants, study design and data collection. done in the univariate framework; Henson & Friston, 2007). Here, we provide only a brief description for clarity. One such alternative is Constrained Principal Component Analysis for fMRI (fMRI-CPCA). fMRI- Participants CPCA extracts functional brain networks from whole- brain Blood Oxygen Level Dependent (BOLD) signal data 33 healthy, right-handed adults (22 females, 11 males) participat- with variance constrained to that predictable from task ed in the study, with a mean age of 27.9 years (SD = 9.0 years). timing, and generates spatial maps, as well as estimates of All participants provided informed consent and reported no prior hemodynamic responses (HDRs) for each combination of history of psychiatric, neurological or pain disorders. Ethical re- subject, task condition and brain network. The technique is view and approval were provided by the Columbia University valuable in that it combines: (1) networks based on multi- Institutional Review Board (Protocol number AAAE3743). variate analyses, which interrogate the intercorrelated struc- Since the data were anonymized and we performed a secondary ture of task-based voxel data without submitting each voxel analysis, no local ethics review was required. to a separate statistical test as in univariate approaches (e.g. In our study, two participants (subjects 11 and 30) were seed-based connectivity techniques, where each voxel is excluded because they received too few trials under each ex- correlated to the seed), (2) networks extracted from BOLD perimental condition (defined below), creating problems for signal constrained to task-timing-related variance, which is the fMRI-CPCA algorithm. This left 31 participants to be useful because task-optimized networks can be more readily analysed. associated with cognitive and behavioural functions by analysing how network HDRs differ between task condi- Thermal Stimulation tions, and (3) data-driven network extraction, meaning that no assumptions about the spatial or temporal properties of To elicit pain, a thermode device was placed on the volar networks are formally defined. Spatial and temporal as- surface of the left forearm (TSA-II Neurosensory Analyzer sumptions are avoided by analyzing all voxels in the brain with a 16-mm Peltier thermode endplate, Medoc Advanced instead of selecting regions-of-interest and using a Finite Medical Systems). Thermal stimuli were delivered at specific Impulse Response (FIR) model of task-evoked HDRs in- temperatures for 12.5 s each, with 3 s of ramp-up, 7.5 s at the stead of assuming a particular HDR shape, respectively. target temperature, and 2 s of ramp-down. Temperature levels Neuroinform (2022) 20:155–172 157 ranged from 40.8 °C to 47.3 °C (study documentation and and the AC-PC plane was oriented horizontally. Slice-time participant results are available at https://openneuro.org/ correction was performed to mitigate the temporal lag in slice datasets/ds000140/versions/00001). acquisition across the 2-s TR, using slice 21 as a reference. Realignment algorithms were applied to counteract displace- fMRI Task ment of voxels due to head movement, and runs that exceeded movement parameter thresholds of 4.5 mm in either z, x, y Participants completed 9 separate functional scanning ses- direction, as well as pitch, yaw or roll, by at least 50 scans, sions. There were 3 types of sessions: “standard” runs, where were removed from the analysis (subject 10, run 5 and 6; pain stimulation was received passively; “regulate-up” runs, subject 2, run 5; subject 4, run 1, 4 and 5). For each partici- where participants were instructed to increase the intensity of pant, functional scans were co-registered to their correspond- pain by cognitive control; and their counterpart, “regulate- ing structural images, and structural T1 scans were segmented down” runs. The regulation manipulation was intended to en- into gray matter, white matter, cerebrospinal fluid, meninges gage supplementary brain systems for pain regulation. For the and skull components. Finally, raw functional data were nor- explicit purposes of our study, we focused on standard runs malized to MNI template space (with a voxel size of 3 × 3 × only. 3 mm) and smoothed with a 6 × 6 × 6 FWHM Gaussian Each standard run began with an 18-s fixation cross pre- kernel. sented on screen, followed by 11 consecutive trials. Each trial For detailed explanations of preprocessing methods, along was 33–41 s long and featured the same progression: 12.5 s of with specific versions of software tools used, refer to supple- thermal stimulation, 4.5–8.5 s (jittered) of pre-rate rest, 11 s of mentary materials. pain rating (completed on screen using a hand-held remote), and 5–9 s (jittered) of post-rate rest. The rating period in- Task-Based Whole-Brain Network Analysis volved two kinds of rating; first, participants decided whether a stimulus was painful or not (this phase lasted 4 s), then Constrained Principal Component Analysis (CPCA) is a sta- participants rated the intensity of their sensation on a Visual tistical technique that combines multivariate least-squares re- Analogue Scale from 0 to 200, where the interval 0–100 rep- gression with principal component analysis (Hunter & resented non-painful warmth, and 100–200 represented the Takane, 2002; Takane & Shibayama, 1991; Takane & intensity of a stimulus perceived as painful. The scale was Hunter, 2001). It can be used to perform whole-brain analyses presented on screen, and participants were instructed as to of fMRI BOLD signal data. When applied to fMRI, it iden- the meaning of each interval prior to scanning. The specific tifies multiple functional networks involved in a task and es- order of temperatures administered throughout each run can timates fluctuations in BOLD signal for each network, over a be found in Woo et al. (2015). For a schematic illustration of specified interval of time. Further statistical tests can be used task design, see Fig. 1. to quantify the interactions between networks, correlational relationships between network activation and behavioural Image Acquisition measures, and the effect of experimental manipulations on the activation of each network. Whole-brain functional images were collected on a 3 T Philips Broadly speaking, fMRI-CPCA involves two steps. First, Achieva TX scanner at Columbia University’s Program for multivariate least-squares multiple regression is used to isolate Imaging in Cognitive Science (PICS). Structural images were variance in BOLD signal that is predictable from the timing of collected with high-resolution T1 spoiled gradient recall im- stimulus presentation, after which the variance is said to be ages (SPGR), which allow for anatomical localization and “constrained” to task timing. This first step is referred to as the warping to standard space. For functional EPI image collec- external analysis. Second, a principal component analysis tion, the following scanning parameters were set: TR = (PCA) is conducted on the constrained portion of the variance 2000 ms, TE = 20 ms, field of view = 224 mm, 64 × 64 ma- in BOLD signal, and the extracted components represent sys- trix, 3 × 3 × 3 mm voxels, 42 interleaved slices, parallel im- tems of functionally interconnected voxels (i.e. functional aging, SENSE factor 1.5. E-Prime software (PST Inc.) was brain networks) related to the task. This step is referred as used to control stimulus presentation and collect behavioural the internal analysis. Importantly, applying PCA after the re- data. gression ensures that the networks identified are based on task-related information only. This is a defining feature of Preprocessing fMRI-CPCA and distinguishes it from other applications of PCA (or ICA) used in fMRI. In fMRI-CPCA, the variance For our analysis, all preprocessing was completed in SPM 12. shared between principal components and task timing is max- Structural and functional scans were reoriented manually, imized, thus avoiding any contamination of the solution by such that the origin was placed on the anterior commissure, variability that is not predictable from task timing. Ultimately, 158 Neuroinform (2022) 20:155–172 Fig. 1 Schematic illustration of task design adapted from Woo et al. pre-rate anticipation period (4.5–8.5 s, jittered), a 4 s rating period to (2015). Every run was preceded by an 18 s fixation cross presented on judge if the stimulus was painful or not, a 7 s pain rating period using a screen. Every trial began with a 12.5 s thermal stimulus, followed by a VAS scale, and a post rate rest period of 5–9 s (jittered) fMRI-CPCA outputs brain activity maps that can be overlaid zero is in all other cells—thus, the G matrix simply defines the on a structural image (for example, in applications like time intervals during which we expect to see task-relevant MRIcron [https://www.nitrc.org/projects/mricron]), as well activations. The number of rows in the G matrix will equal as estimated hemodynamic response shapes (plotted over the number of rows in the Z matrix, but the number of columns post-stimulus time) for each combination of network, is equal to the number of post-stimulus time points (time bins) subject and task condition. The next few paragraphs for which the BOLD signal is to be predicted, multiplied by will elaborate on specific matrices and equations the total number of conditions and the total number of sub- required to implement the analyses. jects. The G matrix is also standardized for each individual In order to perform the external analysis, two matrices must run. We then regress the Z matrix onto the G matrix, first be prepared. The Z matrix (or activation matrix) contains the BOLD data for all runs, with each voxel represented as a Z ¼ GC þ E; single column, and each full-brain scan represented as a single −1 where C =(G ′ G) G ′ Z is a matrix of timepoint- and voxel- row. In the current study, 31 subjects went through nine runs specific regression weights that satisfy the least-squares crite- each, with 209 scans per run. Six runs were removed due to rion. When C is applied to G it provides a matrix of BOLD excessive head movement (see section “2.5. Preprocessing”), signal values predicted from task-timing, Z or GC. E leaving a total of 42,427 rows (full brain scans) and 79,522 scans voxels columns (voxels) in the Z matrix. The mean value for each represents the residual signal (i.e. signal that is not predictable voxel was centered to zero for each run separately, and the from task-timing), which is disregarded in the rest of the anal- variables standardized (such that the standard deviations were ysis. As an additional note, E can be further analysed exactly set to one for each run separately). The G matrix (or design like GC; such an analysis would produce the dominant net- matrix) contains a Finite Impulse Response (FIR) model of the works that are not predictable from task timing, which may be BOLD signal based on stimulus presentation timing; unlike those engaged during off-task periods, or task on processes more conventional models, the FIR model does not impose a that span the whole series of trials but are not specifically predetermined HDR shape on the dataset (which is commonly elicited by the onset of tasks. This type of analysis was not assumed to aid in determining task-relevant activations in carried out here and is beyond the scope of this paper. BOLD signal). Instead, a value of one is placed into cells of The next stage of fMRI-CPCA is the internal analysis, which typically involves application of a principal component G for which the BOLD signal is to be estimated, and a value of Neuroinform (2022) 20:155–172 159 analysis (PCA) to the constrained, task-related signal (GC). did not substantially improve the solution, only marginally in- This identifies correlated structure underlying the voxel data, creasing variance explained and failingtodetectnew andinfor- grouping correlated voxels into components that represent mative regions/networks, insteadfragmentingnetworksprevi- functional brain networks. Importantly, these components will ously identified in the four-component solution. be optimized to be task-related, because GC contains task- After the external and internal analyses are complete, a related variance only. PCA is achieved through singular value final step is applied to produce estimates of HDR shapes as- decomposition of GC: sociated with each network. This is achieved by relating com- ponent scores (in matrix U) back to stimulus presentation timing (coded in G), and computing P such that: UDV ¼ SVDðÞ GC U ¼ GP; 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 where P contains “predictor weights”—these are weights that vectors. In matrix U, columns represent components, and estimate the intensity of each component for the time bins rows represent scans. The values in matrix U are “component specified in G. When plotted over post-stimulus time, predic- scores” and provide an indication or “score” of how important tor weights reveal the unique HDR shapes elicited by each each component is for each scan. In matrix V, columns repre- subject and condition within each network, for the specified sent components, and rows represent voxels. Cells of V can be pffiffiffiffi interval of time. In this study, predictor weights were averaged rescaled by VD= N to obtain “component loadings”—corre- over subjects before plotting. Further, predictor weights were lation coefficients indicating the correlation of task-related averaged over post-stimulus time to compute overall intensity BOLD signal in each voxel with the respective component values for network activation; more detail on this is provided scores. Voxels that are highly correlated with a given compo- below. nent’s “component scores” form the brain regions that define the functional network represented by that component. Preparation of G Notably, rescaling right singular vectors in V allows them to be interpreted as correlations between voxels and networks, The goal of the current study was to determine how the brain while also providing a better approximation of the inputted configures itself when processing pain, and to use the brain matrix GC by incorporating the variance accounted for by networks detected to generate a model of subjective pain per- each network. To visualize a brain activity map for each net- ception. Accordingly, we formatted the G matrix such that work, columns in rescaled V were overlaid on a brain template separate HDR shapes would be produced for high and low in MRIcron and thresholded to display only the voxels with temperature conditions. The division was based on the median the most dominant loadings (e.g. top 10% absolute values). In temperature administered across all trials, including regulation the current study, we orthogonally rotated and rescaled the V runs (the median temperature was 44.3 °C, any stimulus that matrix prior to display, using a varimax solution with 500 was equal to or less than 44.3 was assigned to the “low” iterations (Abdi & Williams, 2010; Bryant & Yarnold, 1995; temperature condition; the rest were “high”). We examined Kaiser, 1958). brain activity during thermal heat portions of the experimental PCA identifies a large number of components, but a select task only, and only included standard runs in the analysis to few can be extracted (the components that account for the least avoid capturing brain systems for cognitive self-regulation amount of variability are considered noise, or brain activity that is over pain perception. The task-relevant time interval (encoded unlikely to be reliable). Various methods for component selec- in G) was defined as the 16 s immediately following thermal tion exist; in this study we used the elbow method. This method stimulus presentation; in this way, the entire duration of the relies on visual inspection of the scree plot of singular values stimulus and 3.5 s thereafter were accounted for. Because each (Cattell, 1966; Cattell & Vogelmann, 1977). When plotted, sin- full-brain fMRI scan was completed in two seconds, HDRs gular values (which are contained in D) produce a line that grad- were estimated for eight post-stimulus time bins. The G matrix ually approaches zero as components account for less and less therefore consisted of 496 columns (2 conditions × 31 subjects variance. In the elbow method, components are selected for ex- × 8 time bins), and 42,427 rows (equal to the number of rows traction by locating the first abrupt increase in variance—relative in the Z matrix). 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 vari- Preparation of Network and Rating Data for Multiple ance (Kodinariya & Makwana, 2013). In this study, 4 compo- Regressions nents were retained and varimax rotated. We attempted addition- al analyses with a greater number of components retained, to For each network, the predictor weights produced in the final ensure the validity of the chosen threshold. Additional networks step of fMRI-CPCA define the unique HDR shapes associated 160 Neuroinform (2022) 20:155–172 with each condition, over the specified 16-s time interval. To sample was maintained between 45% and 55% (non- model pain perception from brain activation, it was preferable inclusive). to compute a single value that would capture the intensity of All brain networks detected by fMRI-CPCA (components the response. In this case, due to exploration of the HDR 1, 2, and 4) were then used to model the pain rating variable, shapes obtained, network activation intensity was estimated Rating∼1 þ Component1 þ Component2 þ Component4; by averaging predictor weights (i.e. estimated BOLD signal) over the entire post-stimulus time interval. This yielded 248 for every sample drawn (305 samples in total). This effective- estimates in total, one for each temperature category for each ly treated the regressors as random rather than fixed effects of the four networks detected by fMRI-CPCA for every (Fox, 2015), and it provided empirical bootstrap distributions participant. for relevant statistics like regression coefficients, R and mod- Pain ratings were subjected to a similar procedure: ratings el significance as determined by F-test, from which estimates associated with each temperature condition were averaged of each metric or model parameter could be obtained. over trials to obtain participant-specific estimates of pain per- Confidence intervals (95%) were calculated non- ception during high- and low-temperature stimuli, yielding 62 parametrically using percentiles (Fox, 2015). estimates in total. Model fit was further evaluated by determining the accura- cy with which fitted values distinguished between pain and Multiple Regressions warmth. This was done by converting the ratings predicted by the model (a continuous variable) into a categorical outcome, Two separate multiple linear regression analyses were con- pain or non-pain, based on the 100-point pain threshold spec- ducted on pain rating and network activation data. ified by the VAS scale. For every model, the predicted binary outcomes were compared to the true state of affairs in order to generate estimates of accuracy (the proportion of total cases Modelling Within-Subject Pain that were correctly classified as either pain or warmth), sensi- tivity (the proportion of pain cases that were correctly classi- The first of these modelled changes in perceived pain as a fied as pain) and specificity (the proportion of non-pain cases function of changes in the intensity of network activations. that were correctly classified as warmth). This provided em- The fundamental goal here was to examine how a change in pirical bootstrap distributions and corresponding estimates rating corresponded with changes in activation intensities be- (with percentile intervals) of model accuracy, sensitivity and tween high and low temperatures. All brain networks detected specificity. All regression analyses were completed in by fMRI-CPCA (component 3 was excluded because it MATLAB R2019a (scripts available from https://github. reflected a movement artifact, see section "3. Results") were com/MatteoDamascelli/Multiple-Functional-Brain- inputted as predictors to explain changes in pain rating: Networks-Related-to-Pain-Perception-Revealed-by-fMRI.). ΔRating∼1 þ ΔComponent1 þ ΔComponent2 þ ΔComponent4: Results To evaluate model fit beyond R , fitted values were plotted against, and correlated with, the response variable using Summary of fMRI-CPCA Output Pearson’s r correlation coefficient. The scree plot of singular values indicated that four compo- Modelling Between-Subject Pain nents should be extracted. Components 1, 2, 3, and 4 accounted for 21.47, 7.26, 4.95, and 3.74% of task-related To investigate the relationship between perceived pain and variance in BOLD signal, respectively. Component images network activation intensity across subjects, we applied a and estimated HDR shapes are displayed in Figs. 2, 3, 4, 5, bootstrap-like regression procedure. Samples of size n =31 along with network activation intensities for each temperature were drawn from the dataset of condition- and participant- category (box plots). specific estimates of pain ratings and network activation in- Component 1 was primarily comprised of a) motor areas, tensities (see section “2.8. Preparation of Network and Rating including the primary motor cortex (M1), supplementary mo- Data for Multiple Regressions”). Each participant contributed tor area (SMA) and cerebellum, b) visual areas, including the one pain rating (and its corresponding network intensities), lateral occipital cortex (LO), and c) the primary somatosenso- selected at random, and every sample was a near-balanced ry cortex (S1). Component 2 featured a frontoparietal activity combination of ratings greater than 100 (i.e. painful) and rat- pattern that included activation peaks in the anterior cingulate ings less than 100 (i.e. warm). The prevalence of pain in the cortex (ACC), dorsolateral prefrontal cortex (dlPFC), anterior Neuroinform (2022) 20:155–172 161 Fig. 2 a-c Component 1. a Three-dimensional rendering of Component 1 distributions of Component 1 BOLD signal across participants, for both (based on the top 10% of component loadings) and estimated HDRs high and low temperature stimuli. BOLD signal was first averaged over associated with this network over the course of one thermal stimulation the entire post-stimulus time interval for each participant and each con- trial. The red bar placed over x-axis tick labels indicates the duration of a dition. The mean is given by ×. c Horizontal cross-sections of Component thermal stimulus (12.5 s). Estimated HDRs were obtained by averaging 1 (only the top 10% of component loadings are shown). Positive loadings the FIR-based predictor weights for each condition level and plotting in red, threshold = 0.17, max = 0.33. No negative loadings. Blue values them as a function of post-stimulus time. Error bars given by standard indicate the MNI coordinate of each slice in the z direction error. HDR = hemodynamic response. b Boxplots illustrating the Fig. 3 a-c Component 2. a Three-dimensional rendering of Component 2 distributions of Component 2 BOLD signal across participants, for both (based on the top 10% of component loadings) and estimated HDRs high and low temperature stimuli. In both cases, BOLD signal was first associated with this network over the course of one thermal stimulation averaged over the entire post-stimulus time interval for each participant trial. The red bar placed over x-axis tick labels indicates the duration of a and each condition. The mean is given by ×. c Horizontal cross-sections thermal stimulus (12.5 s). Estimated HDRs were obtained by averaging of Component 2 (only the top 10% of component loadings are shown). the FIR-based predictor weights for each condition level and plotting Positive loadings in red, threshold = 0.09, max = 0.21. No negative load- them as a function of post-stimulus time. Error bars given by standard ings. Blue values indicate the MNI coordinate of each slice in the z error. HDR = hemodynamic response. b Boxplots illustrating the direction 162 Neuroinform (2022) 20:155–172 Fig. 4 a-c Component 3. a Three-dimensional rendering of Component 3 HDR = hemodynamic response. b Boxplots illustrating the distributions (based on the top 10% of component loadings) and estimated HDRs of Component 3 BOLD signal across participants, for both high and low associated with this network over the course of one thermal stimulation temperature stimuli. In both cases, BOLD signal was first averaged over trial. The red bar placed over x-axis tick labels indicates the duration of a the entire post-stimulus time interval for each participant and each con- thermal stimulus (12.5 s). Estimated HDRs were obtained by averaging dition. The mean is given by ×. c Horizontal cross-sections of Component the FIR-based predictor weights for each condition level and plotting 3 (only the top 10% of component loadings are shown). Negative load- them as a function of post-stimulus time. Blue coloring indicates negative ings in blue, threshold = −0.08, max = −0.16. Positive loadings in red, loadings; graphs should be interpreted as displaying the intensity of de- threshold 0.08, max = 0.09. Blue values indicate the MNI coordinate of activation instead of activation. Error bars given by standard error. each slice in the z direction Fig. 5 a-c Component 4. a Three-dimensional rendering of Component 4 HDR = hemodynamic response. b Boxplots illustrating the distributions (based on the top 10% of component loadings) and estimated HDRs of Component 4 BOLD signal across participants, for both high and low associated with this network over the course of one thermal stimulation temperature stimuli. In both cases, BOLD signal was first averaged over trial. The red bar placed over x-axis tick labels indicates the duration of a the entire post-stimulus time interval for each participant and each con- thermal stimulus (12.5 s). Estimated HDRs were obtained by averaging dition. The mean is given by ×. c Horizontal cross-sections of Component the FIR-based predictor weights for each condition level and plotting 4 (only the top 10% of component loadings are shown). Negative load- them as a function of post-stimulus time. Blue coloring indicates negative ings in blue, threshold = −0.06, max = −0.14. Positive loadings in red, loadings; graphs should be interpreted as displaying the intensity of de- threshold 0.06, max = 0.07. Blue values indicate the MNI coordinate of activation instead of activation. Error bars given by standard error. each slice in the z direction Neuroinform (2022) 20:155–172 163 Table 1 Within-subject pain: estimated regression coefficients and and posterior insula (aIns and pIns), and thalamus. related statistics Component 3 was limited to the outer edge of the cortex (specifically the frontal and occipital poles) and the longitu- Predictor b (S.E.) β t p value dinal fissure and was mostly composed of negative loadings (Intercept) 40.78 (7.19) 5.67 0.000 (i.e. it became deactivated during stimulation). This particu- Component 1 -138.94 (79.41) −0.29 −1.75 0.092 lar configuration was biologically untenable and resembled Component 2 201.39 (54.54) 0.65 3.69 0.001 no established networks; it was most likely summarizing Component 4 112.07 (40.77) 0.48 2.75 0.011 head movement that was coordinated with the application of thermal pain. Component 4 was characterized by deactivations b = unstandardized regression coefficient, β = standardized regression co- in areas conventionally associated with the default mode net- efficient, S.E. = standard error, p-values are two-tailed work, including the posterior cingulate cortex (PCC), medial prefrontal cortex (mPFC), precuneuous and angular gyrus consisted primarily of negative loadings (i.e. its intensity (AnG). Detailed anatomical descriptions for all networks are values represented deactivation, not activation, intensity). found in supplementary information (supplementary tables 1–3). To further evaluate model fit, changes in pain rating were plotted against changes predicted by the model in Fig. 6.The two variables were significantly correlated (r = 0.630, Regressions p < .001). Importantly, this model shows that the networks identified by fMRI-CPCA, as a whole, capture variations in To relate these networks back to pain perception, we modelled pain perception at the within-subject level. 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 rat- Between-Subject Pain ings and network activation intensities for each temperature category (i.e. high and low) within each participant. Bootstrapped regression models of pain ratings, with all func- Descriptive statistics (means and standard errors) for these tional networks inputted as predictors, explained 28.6% of the data are found in supplementary table 4. variance in pain rating data on average (R = 0.286, CI = [0.079, 0.475]), or 20.7% when adjusted for the num- 95% ber of predictors (R = 0.207). F-tests for the variance Within-Subject Pain adjusted accounted for by each model revealed that 65.9% of the time, For within-subject pain, the model included temperature- models were significant at the .05 level with a median p value dependent changes in network activation intensity for all func- of .023 (see Fig. 7). tional networks identified by fMRI-CPCA (components 1, 2 The accuracy of predicted pain ratings was evaluated by and 4), and explained 39.7% of the variance in temperature- averaging RMSE across all re-sampled models (RMSE = dependent changes in pain rating (R =0.397; F(3, 27) = 5.9, p 37.24; Fig. 7). Estimated regression coefficients, their stan- = .003), or 32.9% when adjusted for the number of predictors dard errors and confidence intervals are given in Table 2; (R = 0.329). This indicates that changes in pain ratings components 2 and 4 were the only significant predictors of adjusted are predictable from changes in BOLD signal in the functional pain rating (Component 2: β = 0.42, CI = [0.12, 0.64]; 95% networks identified, according to: Component 4: β =0.26, CI = [0.09, 0.44]). Further, stan- 95% dardized coefficients showed that Component 2 made the change in rating∼−138:94ðÞ C1 BOLD signal intensity change most important contribution to the model, and its activation þ 201:39ðÞ C2 BOLD signal intensity changeþ intensity was positively associated with pain ratings. By con- trast, Component 4 deactivation intensity was positively asso- 112:07ðÞ C4 BOLD signal intensity changeþ 40:78: ciated with pain ratings. Figure 8 provides a schematic sum- The accuracy of predicted scores was evaluated by taking mary of these relationships between networks and pain per- the standard deviation of the residuals or the Root-mean- ception. Overall, this model shows that—as a whole—the square error (RMSE), which was 18.27. As shown in networks delineated by fMRI-CPCA are sensitive to be- Table 1, components 2 and 4 were the only statistically sig- tween-subject variability in pain perception. nificant contributors (Component 2: β = 0.654, p =.001; To evaluate the model’s ability to differentiate be- Component 4: β = 0.482, p = .011). Also of note, tween painful and non-painful states, we converted true Component 2 predicted increases in pain based on increases ratings and predicted ratings into binary categories (i.e. in its activation, whereas Component 4 predicted increases in pain or non-pain, as described in method) and calculat- pain based on increases in its deactivation, given that it ed accuracy, sensitivity and specificity of classification 164 Neuroinform (2022) 20:155–172 Fig. 6 a-d Assessing model fit for within-subject pain. Temperature- model. d True change in pain ratings plotted against change in ratings dependent change in pain ratings plotted against change in BOLD signal predicted by the model. The strength and significance of the relationship for Component 1 (a), Component 2 (b), and Component 4 (c); true is given by Pearson r = 0.630, p < .001 changes are shown alongside predictions made by the linear regression for every resampled model; Supplementary Figure 1 predicted ratings tend to better approximate ratings be- demonstrates this procedure for one bootstrap sample. low the pain threshold than ratings above it, such that Empirical bootstrap distributions for accuracy, sensitivi- when ratings are converted to the binary variable “pain” ty and specificity are shown in Fig. 9. Estimates (ob- or “non-pain”, ratings above the threshold are more of- tained by averaging) were 68.83%, 59.17% and 77.10%, ten miscategorized than ratings below it. respectively. Confidence intervals show that only accu- racy and specificity were significantly higher than chance level (50%). This indicates that the overall ac- Discussion curacy of the model is driven by its specificity, or its ability to correctly identify non-pain (more precisely, In this study, distinct functional connectivity networks for specificity is the fraction of pain ratings below 100 pain were revealed by fMRI-CPCA. The networks encompass correctly classified as “non-pain”). It appears that the a variety of brain regions consistently active in response to Fig. 7 Assessing model fit for between-subject pain. Histograms, kernel 2 2 three figures: R ¼ 0.286, CI = [0.079, 0.475]; R = 0.207, 95% adj: density estimates, and average bootstrap estimates for R-squared, adjust- CI =[−0.023, 0.417]; RMSE = 37.238, CI = [31.247, 43.404]. 95% 95% ed R-squared, Root-mean-square Error (RMSE) and model significance For model significance: 65.9% of models were significant at the .05 level, afforded by F-test (determines if a model fits significantly better than one p = .023 median based on a constant term only). Means and percentile intervals for the first Neuroinform (2022) 20:155–172 165 Table 2 Between-subject pain: Unstandardized coefficient Standardized coefficient (β) estimated regression coefficients and related statistics Predictor Average Bootstrap Percentile CI (95%) Average Bootstrap Percentile CI (95%) Estimate (S.E.) Estimate (S.E.) (Intercept) 73.67 (16.81) [36.14, 103.13] Component 1 −48.95 (78.65) [−228.15, 87.88] −0.07 (0.11) [−0.30, 0.13] Component 2 172.26 (47.66) [52.76, 255.13] 0.42 (0.12) [0.12, 0.64] Component 4 76.24 (27.04) [24.54, 133.80] 0.26 (0.09) [0.09, 0.44] S.E. = standard error, i.e. the standard deviation of the corresponding bootstrap distribution pain, including MI, SMA, cerebellum and SI (Component 1), functional networks. Although the specific parcellation observed the ACC, insular cortex, and thalamus (Component 2), and here is unique, it is largely congruent with current perspectives mPFC, hippocampus, para-hippocampus and precuneus on pain-related networks. In particular, evidence from PET and (Component 4) (Apkarian et al., 2005; Atlas et al., 2014; fMRI suggests that pain-activated regions are segregated into at Schweinhardt & Bushnell, 2010). Within participants, chang- least four distinct sub-networks: a sensory network for stimulus es in perceived intensity related to low and high temperatures localization and intensity coding (Davis & Moayedi, 2013; were associated with the magnitude of change in BOLD Hofbauer et al., 2001;Peyronetal., 1999), an affective network across networks. While falling short of accurate classification, for generating the aversive, unpleasant quality of a stimulus the magnitude of BOLD activation in functional networks was (Davis & Moayedi, 2013;Peyron et al., 2000;Wilcoxetal., significantly associated with pain intensity between partici- 2015), a cognitive network for attending to, anticipating and pants. Future development of fMRI-CPCA in the context of remembering the stimulus (Davis & Moayedi, 2013;Peyron pain is warranted to further explore the brain in pain. et al., 1999;Wilcoxetal., 2015), and a network of motor regions Over and above capturing the activation of known pain re- for pain avoidance (Davis & Moayedi, 2013;Wilcoxet al., gions, fMRI-CPCA integrated these brain regions into multiple 2015). Fig. 8 a-b Component contributions to pain. a Schematic depiction of each component’s contribution to pain ratings, according to bootstrap estimates of standardized regression coefficients (beta weights). Lines are proportional to the strength of their relation to pain. Component 4 is characterized by deactivation instead of activation; thus, its β value captures the strength of the relationship between Component 4 deactivation and pain perception. b Histograms, kernel density estimates, and average bootstrap estimates for unstandardized regression coefficients, with means and CI bounds. Component number indicated in the top left corner 166 Neuroinform (2022) 20:155–172 Fig. 9 Classification performance. Histograms, kernel density estimates, and average bootstrap estimates of model accuracy (Mean = 68.83, acc. CI = [54.84, 83.87]), 95% sensitivity (Mean = 59.17, sens. CI = [33.33, 80.63]) and 95% specificity (Mean = 77.10, spec. CI = [57.14, 93.75]) 95% Component 1 (Sensorimotor Response) observed because of screen-related cues that coincided with stimulus presentation. Component 1, with prominent activation peaks in MI, SMA, and cerebellum, is most aptly described as a sensory and Component 2 (Attentional Pain Network) motor network. In the context of thermal stimulation, sensa- tion and motor output may be related to an instinctive In agreement with previous studies, Component 2 incorporat- flexing or bracing, or a desire to move, in response to in- ed a large number of regions involved in pain, and combined tense stimuli (Davis & Moayedi, 2013; Davis et al., 2002). sensory, affective and cognitive sub-networks (Davis & Hemodynamic response shapes (HDRs) for Component 1 Moayedi, 2013;Wilcox et al., 2015). For example, the most showed that activation was in fact exclusive to higher inten- prominent activation peaks were found in SII and posterior sity stimuli (temperatures above 44.3 °C). Component 1 also insula (pIns; sensory-discriminative regions), dACC, aIns, included prominent activations in SI, which, as a key corti- and thalamus (affective-motivational regions), and dlPFC cal aspect of the lateral nociceptive system, is one of the first and IPL (cognitive-evaluative regions; Peyron et al., 1999; recipients of ascending pain signals through the Wilcox et al., 2015). Based on this, Component 2 could reflect spinothalamic tract (Davis & Moayedi, 2013;Fomberstein a unification of sensory, affective and cognitive processes et al., 2013; Vierck et al., 2013; Yam et al., 2018). SI’s (Melzack & Casey, 1968) into a coordinated pain response. inclusion in Component 1 suggests that, during thermal The blending of sub-networks is likely facilitated by their stimulation, motor processes are prioritized and closely co- inter-connectivity at rest, provided by common nodes in ACC ordinated with sensory-discriminative functions (e.g. deter- and aIns that serve as relay sites between sub-networks mination of stimulus location and intensity). In theory, such (Wilcox et al., 2015). Importantly, the ACC and aIns are en- close communication would be necessary for an effective gaged in non-specific “salience detection”, where stimuli are pain avoidance response when stimulus intensity reaches selected based on their behavioural relevance, and attentional noxious levels (Postorino et al., 2017). This plausible role systems are primed to enable an effective response (Legrain of Component 1 in generating pain-induced motor com- et al., 2011; Menon, 2015). Such a “salience network” re- mands remains to be further explored; follow-up studies ceives axonal projections from sensory areas like the pIns, would benefit from monitoring physical movements in con- which are thought to provide the aIns with incoming sensory junction with other variables, allowing for the precise rela- information (Menon & Uddin, 2010). The pattern of activa- tionships between Component 1 activation intensity, stimu- tion observed in Component 2 captures both attentional sys- lus intensity (e.g. temperature), motion, and pain perception tems (i.e. the cognitive sub-network of pain) and sensory- to be determined. discriminative elements like SII and pIns. Thus, Component A novel observation from fMRI-CPCA is the temporal 2 may represent a salience network-mediated response to overlap between visual areas and sensory-motor coupling, ev- salience—in this case thermal stimulation—or more precisely idenced in Component 1. Their detection is likely an idiosyn- a sequential activation of sub-networks, i.e. sensory systems cratic capture of fMRI-CPCA, which avoids using regions-of- activate the salience network, which then activates cognitive interest to spatially constrain the analysis. In fact, the anatomy systems for sustained attention. The directionality of sub- of Component 1 replicates previous applications of fMRI- network relations is a matter of speculation, but it presents CPCA in other domains—specifically, it resembles a network an interesting question for future investigation. Additionally, consistently associated with sensorimotor response processes, the putative attentional function of Component 2 may be fur- featuring activations in lateralized MI, SMA, SI, cerebellum, ther explored by analyzing its pain-induced response during and visual areas including the lateral occipital cortex (LO; experimental manipulations of attentional demand or stimulus Goghari et al., 2017; Larivière et al., 2017; Metzak et al., salience; a larger effect of attention on network response than 2011). In this case, sensorimotor-visual coupling was likely stimulus temperature would suggest an attentional role. Neuroinform (2022) 20:155–172 167 Component 4 (Default-Mode Network) alterations to the DMN. This possibility requires further inves- tigation and presents an important research objective due to its The tendency for brain areas to become deactivated during a implications for chronic pain treatment. 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, Estimating Pain Within and Between Participants research has emerged documenting the overall functional con- tributions of the default-mode network (DMN) to human be- Among intended applications of neuroimaging in the field of havior, including its relevance to mind-wandering, self- pain is the development of models to accurately classify an referential thought, mentalizing and semantic processing individual in pain. Previous attempts of this nature have (Andrews-Hanna, 2012; Andrews-Hanna et al., 2014; adopted multivariate pattern analysis (MVPA; Haynes, Christoff et al., 2009). 2015; van der Miesen et al., 2019). In brief, MVPA uses Component 4 was comprised of deactivations in regions machine learning algorithms to model behavioural responses conventionally associated with the DMN, including the (either ordinal or continuous variables) as a function of mul- PCC, the AnG, and the amPFC. Such a deactivation departs tiple voxels (or “features”) considered simultaneously from the proposed sub-network scheme discussed above (i.e. (Moayedi et al., 2018; van der Miesen et al., 2019); predic- sensory, affective, cognitive, and motor sub-networks of pain; tions or classifications of mental states are then generated on Davis & Moayedi, 2013; Wilcox et al., 2015). However, the independent “testing” data based on model parameters learned DMN has also been implicated in pain and so its detection in the “training” set (Rosa & Seymour, 2014). In one notable here is not entirely unexpected. In chronic pain disorders, for study applying MVPA, a networkofregressionweights example, the DMN shows a number of anatomical-functional distributed over pain regions (the “neurologic pain sig- alterations, including fragmentation between frontal and pos- nature” or NPS) tracked physical pain intensity between terior regions (Baliki et al., 2014), and strengthening of func- individuals (Wager et al., 2013; Woo et al., 2015). tional connections to aIns (Baliki et al., 2014;Loggiaet al., Perhaps even more remarkable is that physical pain 2013). In healthy individuals, heat-induced deactivations in could be accurately distinguished from other types of several DMN regions have been reported (Kong et al., pain (e.g., social; Wager et al., 2013). 2010), while some regions, like the hippocampus and In this study, regression models provided some insight into precuneus, also predict pain ratings (in addition to stimulus the capacity of networks detected through fMRI-CPCA to be intensity) by the magnitude of their deactivation (Atlas et al., used for pain prediction, as components 1, 2 and 4 were sig- 2014). nificantly associated with pain ratings both within and be- As others have argued, pain-induced deactivations in the tween participants. Importantly, this was not a predictive mod- DMN may be part of an attentional response to pain (Kucyi el (networks were used to model in-sample ratings with no et al., 2013; Kucyi & Davis, 2015), where the DMN sup- predictions generated on new or held-out data), and the find- presses as attentional networks (e.g. Component 2) engage. ings should not be interpreted as direct evidence of prediction This type of antagonistic relationship between the DMN and ability. However, networks did show potential to be used in attentional networks has been documented extensively outside predictive analyses given that in-sample estimation was mod- of pain imaging, along with the DMN’s “task-negative” ten- erately accurate, and, importantly, results were achieved with- dencies (Anticevic et al., 2012; Peng et al., 2018). Future out any a priori selection of brain regions, reflecting a distinct research would benefit from an analysis of DMN response advantage of fMRI-CPCA compared to other approaches. to pain in the context of attentional manipulations. Of all networks, Component 2 was most strongly related to Alternatively, attention levels during a stimulus could be mon- pain perception; the relationship was positive and consistently itored to allow for an analysis of the relationships between accounted for the largest proportion of within- and between- DMN deactivation, DMN-Component 2 antagonism, pain subject variability in pain. The value of Component 2 for perception and attention. predicting pain is intuitive, insofar as brain regions included Also of note, several DMN regions, including the mPFC, in this network represent sensory, affective, and cognitive di- hippocampus, and precuneus, have been associated with the mensions of pain (Melzack & Casey, 1968). The DMN was regulation of pain (Goffaux et al., 2014; Schweinhardt & also important for pain estimation, with the magnitude of its Bushnell, 2010). Their involvement implies a potential role deactivations being significantly related to perceived pain in- of the DMN, which might accomplish regulation by tensity, both within and between participants. The relation- interacting with the periaqueductal gray (PAG)—part of a ships of both networks to pain are corroborated by previous descending pathway for pain control—through the mPFC work that has identified several Component 2 regions— (Kucyi et al., 2013). Thus, chronic pain disorders may be including SII, aIns, dACC, left cerebellum, and IPL—and related, in part, to deficits in pain regulation caused by DMN regions—including hippocampus and precuneus—as 168 Neuroinform (2022) 20:155–172 explicit mediators of pain (i.e. they mediate the relationship involved in pain perception, without relying on prior assump- between stimulus intensity and pain rating; Atlas et al., 2014). tions about relevant spatial or temporal response patterns. The intensity of activation in Component 1 was unrelated fMRI-CPCA thus provides an opportunity to select to the intensity of perceived pain, mirroring the behaviour of connectivity-based features (Rosa & Seymour, 2014)that SI itself, which codes pain information primarily in terms of are unbiased, data-driven and task-related. sensory-discriminative attributes (Moulton et al., 2012). This As a final point, results from multiple regressions are not aspect of Component 1 (i.e. its independence from pain per- only relevant to pain prediction, but also reflect on network ception) is corroborated by mediation analyses that demon- functions proposed earlier, specifically the roles of strate a preference of sensory cortex and cerebellum to stimu- Component 2 and the DMN as attention networks. In the lus intensity over pain report (Atlas et al., 2014), and implies regressions, Component 2 and the DMN displayed opposite that motor systems are mobilized in accordance with stimulus relationships to pain; higher pain was associated with greater properties only; the perception of pain occurs elsewhere, and activation in Component 2 but greater suppression in the the intensity of motor commands is, on its own, an unreliable DMN, both within and between participants. This is an exten- proxy for the intensity of that perception. sion of the pattern shown by estimated HDRs, where Despite significant associations, when converted into a Component 2 became active during stimulation while the classifier the model discriminated between pain and warmth DMN became suppressed. Together, these findings suggest with an accuracy of only 68.83%. While significantly greater that Component 2 and the DMN assume an antagonistic con- than chance, sensitivity and specificity were low (estimated at figuration during pain, and that greater antagonism (i.e. great- 59.17% and 79.10%, respectively). Still, comparisons be- er separation in terms of activation) equates to a heightened tween components 1, 2 and 4 and the existing “neurological perception of pain. pain signature” (NPS) reveal a high degree of overlap. Based on neuroimaging literature, this antagonism is likely Common regions include aIns, pIns, supramarginal gyrus, indicative of an ongoing attentional response. Component 2 thalamus, and IPL. Further, the NPS included negative pre- included known salience network hubs in ACC and aIns, as dictive weights in regions that were deactivated in Component well as cognitive pain regions associated with attention to 4, including PCC, precuneus and mPFC (Wager et al., 2013). pain, and the DMN’s role in attention has been well-docu- These anatomical similarities raise the possibility that accurate mented. For example, the DMN tends to form anticorrelated predictions of pain could be generated from components 1, 2 relationships with frontoparietal attention networks during and 4 if specific regional activations (compared to an overall cognitively demanding tasks (Dixon et al., 2017; Menon, estimate of activation in the entire network) were accounted 2015; Sridharan et al., 2008), with greater deactivation for using MVPA (Allefeld & Haynes, 2015). By avoiding predicting improved task performance (Anticevic et al., spatial averaging, MVPA accounts for signal non- 2012). Furthermore, attention deficits are generally associated uniformities between voxels, and exploits these differences with increased DMN activation (Bonnelle et al., 2011; in response signal as a source of predictive information Weissman et al., 2006; Danckert & Merrifield, 2018). In the (Hebart & Baker, 2018). context of pain, DMN deactivation is especially pronounced Crucially, the predictive potential shown by components when participants report attending to pain, and less so when indicates that fMRI-CPCA may provide a useful tool for de- participants mind-wander away from pain (Kucyi et al., 2013). termining appropriate anatomical targets for MVPA. This is Thus, the deactivation of DMN observed here likely signifies important because a critical step in the MVPA framework is attention to pain. The simultaneous activation of Component the selection of “features” with which to train the machine 2—which included several regions known to be involved in learning algorithm (Rosa & Seymour, 2014). Features are typ- attention—mirrors the stereo-typical antagonism between ically a subset of voxels, whose activations will be related to DMN and frontoparietal networks that underlies attention the behavioural response by the algorithm (Allefeld & (Anticevic et al., 2012). In sum, these networks appear to Haynes, 2015), and are selected from a region- or regions- contribute to pain perception by working together, in an of-interest (based on prior knowledge) or from the entire brain anticorrelated fashion, as part of an attentional response pro- using dimensionality reduction techniques like PCA (van der cess; the greater the attention, the greater the antagonism be- Miesen et al., 2019). Restricting the analysis to relevant re- tween networks and the greater the pain intensity. gions is important to mitigate the problem of features exceed- ing the number of observations, which may lead to overfitted Technical Considerations models and interpretive challenges (van der Miesen et al., 2019). In the case of the NPS, features were selected a priori Based on the literature discussed in sections above, it is pos- from a collection of well-established pain regions (Wager sible to infer the functionality of each network. However, et al., 2013). By contrast, fMRI-CPCA would allow features these inferences are speculative and are not necessarily vali- to be selected from the predominant functional networks dated by any direct experimental evidence obtained here; Neuroinform (2022) 20:155–172 169 instead they rely on prior notions about the functional contri- In theory, the regression model obtained here is generaliz- butions of regions or networks detected. Importantly, the able to new and independent data. The basic procedure would fMRI-CPCA framework provides an opportunity to more ro- involve application of the regression coefficients obtained to bustly characterize network function during a task. This is the network activations estimated in a new individual to gen- done by comparing the HDRs estimated for each network erate a predicted pain rating. This would require first obtaining across task conditions to determine how different combina- an individual’s activation data during a thermal pain task, tions of independent variables impact network behaviour. analysing their brain activity using fMRI-CPCA, and “classi- Statistical comparisons can be made using repeated- fying” the networks elucidated through in-house programs measures ANOVAs, with within-subject factors given by time recently developed to determine which of the new individual’s and independent variables of interest (e.g. temperature level in networks most closely match with the networks that inform this study). By carefully manipulating experimental condi- the current model (Percival et al., 2020). The HDR shapes tions, cognitive processes can be dissociated from each other, associated with the correct networks would have to be aver- and by interpreting main and interaction effects of factors on aged across an appropriate time interval (or an equivalent time HDRs, networks can be related to specific aspects of cognition interval to the current study) to generate estimates of network operationalized by task conditions. Comparisons can also be activation intensity, and the regression model obtained here made between populations of interest by adding between- would then be applied to network activation intensities to subject factors that define group membership. In this way, generate a predicted pain rating. The classification procedure network alterations or deficits associated with diagnostic referred to above has been utilized previously and involves categories—such as chronic pain disorders—can be correlating the loadings of networks obtained with the load- investigated. ings of “template” images of networks, across a set of charac- It should be noted that the HDRs estimated by fMRI-CPCA teristic slices that define the individuality of a network. Each are well-suited to making inferences about cognitive function; network identified in a new individual would have to be cor- this is because fMRI-CPCA uses Finite Impulse Response related with the template images of networks obtained in this (FIR)-basis sets to encode brain activity associated with task- study to determine the strongest matches. Ultimately, this timing, which are essentially dummy regressors for stimulus pre- classification procedure would aid in selecting networks sentation timing that make no assumptions about the shape of the whose activations (averaged over post-stimulus time) would expected response. For this reason, the technique detects re- then be subjected to the regression model in order to generate sponses (and by extension, functional networks) elicited by cog- apain prediction. nitive processes that may go unnoticed in more traditional anal- To address this first limitation (lack of model validation), yses, where the expected response is produced by convolving future research may use larger datasets to re-conduct the cur- rent study with the addition of a validation protocol, or test the stimulus functions with canonical hemodynamic response func- tions (Henson & Friston, 2007;Hensonetal., 2001; Lindquist, current model in new and independent data according to the 2008; Lindquist et al., 2009). Detailed analysis of HDR shapes procedure outlined above. That said, predictions based on evoked in components 1, 2 and 4, under different experimental these networks are likely to be improved if signal differences conditions, is therefore warranted to achieve a robust determina- within sub-networks and regions of components are tion of network function. accounted for by using pattern-based analyses like MVPA, instead of constructing models based on a whole-network in- Limitations dex of activation (i.e. estimated HDRs for entire networks). A second limitation is that we included stimuli not rated as Our study has a number of limitations to consider. First, we painful (based on the 100-point pain threshold specified by the did not include a protocol for model validation when evaluat- VAS scale) in both network extraction via fMRI-CPCA and ing pain predictions and classifications made with multiple regression models of pain ratings. For network extraction, this linear regression; the ability of the model to predict or classify means that networks delineated were composed of voxels that pain in independent samples therefore remains unverified. remained functionally-connected across non-painful and pain- Validation techniques—including cross-validation, hold-out ful stimulation; in this way, any voxels that became incorpo- validation, or bootstrapping—are common practice in rated into the networks—or any new networks that were decoding analyses to ensure that models generalize to out- formed—during painful stimuli only were potentially missed of-sample data (Kohavi, 1995; van der Miesen et al., 2019). by the analysis. For regression models, it raises the possibility We did not apply these here because of properties of the data that networks were related to warmth more so than pain per- (primarily its small sample size of 30), which made a conven- ception. This would be the case if model-based predictions of tional approach challenging (e.g. some subjects never reported ratings below the pain threshold (i.e. warmth) were consistent- a stimulus as painful). Future research is needed to formally ly better than those of ratings above the pain threshold (i.e. pain). validate the pain predictive value of these networks. 170 Neuroinform (2022) 20:155–172 Supplementary Information The online version contains supplementary Conclusion material available at https://doi.org/10.1007/s12021-021-09527-6. Overall, this study has contributed to neuroimaging research Code Availability Analyses were completed with software and custom on pain by elucidating three functional networks evoked by code available and freely accessible online. Specific links are provided in the Information Sharing Statement. thermal stimulation: a sensorimotor response network for im- mediate pain avoidance (Component 1), a frontoparietal atten- Author Contribution Matteo Damascelli: data curation, formal analysis, tion network mobilized by salience detection processes funding acquisition, investigation, methodology, software, visualization, (Component 2), and the default-mode network (Component writing – original draft. Todd S. Woodward: conceptualization, investi- 4). Of these, attention and default-mode networks were related gation, methodology, resources, software, supervision, writing – review & editing. Nicole Sanford: software. Hafsa B. Zahid: software. Ryan Lim: to pain perception both within and between participants. From methodology, software. Alexander Scott: conceptualization, funding ac- a purely technical perspective, this study validates fMRI- quisition, investigation, writing – review & editing. John K. Kramer: CPCA within the domain of pain research for the first time, conceptualization, funding acquisition, investigation, methodology, re- highlighting advantages compared to existing approaches, in- sources, supervision, writing – review & editing. cluding that the parcellation of multiple task-related networks Funding This research was partially funded by the Natural Sciences and is accomplished without a priori selection of regions-of- Engineering Research Council (NSERC; grant numbers 5456 and interest (i.e. no assumptions about spatial properties of net- RGPAS-2017-507820). works). Moreover, fMRI-CPCA does not rely on models that assume specific HDR shapes to identify task-related activity; Data Availability Data are available on a public repository online, and are instead, HDRs are predicted using FIR basis functions, which freely accessible. A link can be found in the Information Sharing Statement. 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 Declarations may be unaccounted for in conventional analyses. Ethics Approval For the original data, ethical review and approval were More generally, the findings obtained provide a foundation provided by the Columbia University Institutional Review Board from which to further investigate these networks, their proposed (Protocol number AAAE3743). Since these data were anonymized and functions and their pain predictive value. The networks identified we performed a secondary analysis, no local ethics review was required. (especially the attention and default-mode networks) may have implications for pain treatments, if they can be targeted success- Consent to Participate All participants provided informed consent. fully with strategies based on neuromodulation (Alo & Competing Interests The authors declare no competing interests. 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 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adap- from these networks and potentially refined by applying tation, distribution and reproduction in any medium or format, as long as MVPA within network boundaries. Patterns delineated through you give appropriate credit to the original author(s) and the source, pro- MVPA may ultimately serve as objective measures of pain, vide a link to the Creative Commons licence, and indicate if changes were which are of crucial importance to effective pain management made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a in patients unable to self-report their pain. credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Information Sharing Statement The thermal stimulation fMRI data that support the findings of References this study were originally published by Woo et al. (2015)and are available from openneuro.org, https://openneuro.org/datasets/ Abdi, H., & Williams, L. J. (2010). Principal component analysis. Wiley ds000140/versions/00001. Accession number ds000140. Interdisciplinary Reviews: Computational Statistics, 2(4), 433–459. The fMRI-CPCA program is available from https://www. Allefeld, C., & Haynes, J. D. (2015). Multi-voxel pattern analysis. In A. nitrc.org/projects/fmricpca; codes are implemented in W. Toga (Ed.), Brain mapping: An encyclopedic reference (pp. 641–646). Academic Press: Elsevier. MATLAB using a graphical user interface. Custom code for Alo, K. M., & Holsheimer, J. (2002). New trends in neuromodulation for within- and between-subject regression models can be accessed the management of neuropathic pain. Neurosurgery, 50(4), 690– from GitHub, https://github.com/MatteoDamascelli/Multiple- Functional-Brain-Networks-Related-to-Pain-Perception- Andrews-Hanna, J. R. (2012). The brain’s default network and its adap- tive role in internal mentation. Neuroscientist, 18(3), 251–270. Revealed-by-fMRI. Neuroinform (2022) 20:155–172 171 Andrews-Hanna, J. R., Smallwood, J., & Spreng, R. N. (2014). The Goghari, V. M., Sanford, N., Spilka, M. J., & Woodward, T. S. (2017). Task-related functional connectivity analysis of emotion discrimina- default network and self-generated thought: Component processes, dynamic control, and clinical relevance. Annals of the New York tion in a family study of schizophrenia. Schizophrenia Bulletin, Academy of Sciences, 1316(1), 29–52. 43(6), 1348–1362. Anticevic, A., Cole, M. W., Murray, J. D., Corlett, P. R., Wang, X. J., & Gorgolewski, C. (2018). Distinct brain systems mediate the effects of Krystal, J. H. (2012). The role of default network deactivation in nociceptive input and self-regulation on pain (00001) [data set]. cognition and disease. Trends in Cognitive Sciences, 16(12), 584– Retrieved from https://openneuro.org/datasets/ds000140/versions/ 00001. AccessedJune2018 Apkarian, A. V., Bushnell, M. C., Treede, R. D., & Zubieta, J. K. (2005). Haynes, J. D. (2015). A primer on pattern-based approaches to fMRI: Human brain mechanisms of pain perception and regulation in Principles, pitfalls, and perspectives. Neuron, 87(2), 257–270. health and disease. European journal of pain, 9(4), 463–484. Hebart,M.N.,&Baker,C.I.(2018).Deconstructingmultivariate Atlas, L. Y., Lindquist, M. A., Bolger, N., & Wager, T. D. (2014). Brain decoding for the study of brain function. Neuroimage, 180,4–18. mediators of the effects of noxious heat on pain. PAIN®, 155(8), Henson, R., & Friston, K. (2007). Convolution models for fMRI. 1632–1648. Statistical parametric mapping: The analysis of functional brain Baliki, M. N., Mansour, A. R., Baria, A. T., & Apkarian, A. V. (2014). images,178–192. Functional reorganization of the default mode network across chron- Henson, R., Rugg, M. D., & Friston, K. J. (2001). The choice of basis ic pain conditions. PLoS ONE, 9(9), e106133. functions in event-related fMRI. NeuroImage, 13(6), 149–149. Bonnelle, V., Leech, R., Kinnunen, K. M., Ham, T. E., Beckmann, C. F., Hofbauer, R. K., Rainville, P., Duncan, G. H., & Bushnell, M. C. (2001). De Boissezon, X., et al. (2011). Default mode network connectivity Cortical representation of the sensory dimension of pain. Journal of predicts sustained attention deficits after traumatic brain injury. The Neurophysiology, 86(1), 402–411. Journal of Neuroscience, 31(38), 13442–13451. Hunter, M. A., & Takane, Y. (2002). Constrained principal component Bryant, F. B., & Yarnold, P. R. (1995). Principal-components analysis analysis: Various applications. Journal of Educational and and exploratory and confirmatory factor analysis. In L. G. Grimm & Behavioral Statistics, 27(2), 105–145. P. R. Yarnold (Eds.), Reading and understanding multivariate Iannetti, G. D., & Mouraux, A. (2010). From the neuromatrix to the pain statistics (p. 99–136). American Psychological Association. matrix (and back). Experimental Brain Research, 205(1), 1–12. Cattell, R. B. (1966). The scree test for the number of factors. Kaiser, H. F. (1958). The varimax criterion for analytic rotation in factor Multivariate Behavioral Research, 1(2), 245–276. analysis. Psychometrika, 23(3), 187–200. Cattell, R. B., & Vogelmann, S. (1977). A comprehensive trial of the scree and KG criteria for determining the number of factors. Kodinariya, T. M., & Makwana, P. R. (2013). Review on determining Multivariate Behavioral Research, 12(3), 289–325. number of cluster in K-means clustering. International Journal, Chapin, H., Bagarinao, E., & Mackey, S. (2012). Real-time fMRI applied 1(6), 90–95. to pain management. Neuroscience Letters, 520(2), 174–181. Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy Christoff, K., Gordon, A. M., Smallwood, J., Smith, R., & Schooler, J. W. estimation and model selection. International Joint Conference on (2009). Experience sampling during fMRI reveals default network Artificial Intelligence, 14(2), 1137–1145. and executive system contributions to mind wandering. Proceedings Kong, J., Loggia, M. L., Zyloney, C., Tu, P., LaViolette, P., & Gollub, R. of the National Academy of Sciences, 106(21), 8719–8724. L. (2010). Exploring the brain in pain: Activations, deactivations Danckert, J., & Merrifield, C. (2018). Boredom, sustained attention and and their relation. Pain, 148(2), 257–267. the default mode network. Experimental Brain Research, 236(9), Kucyi, A., & Davis, K. D. (2015). The dynamic pain connectome. Trends 2507–2518. in neurosciences, 38(2), 86–95. Davis, K. D. (2011). Neuroimaging of pain: What does it tell us? Current Kucyi, A., Salomons, T. V., & Davis, K. D. (2013). Mind wandering Opinion in Supportive and Palliative Care, 5(2), 116–121. away from pain dynamically engages antinociceptive and default Davis, K. D., & Moayedi, M. (2013). Central mechanisms of pain re- mode brain networks. Proceedings of the National Academy of vealed through functional and structural MRI. Journal of Sciences, 110(46), 18692–18697. Neuroimmune Pharmacology, 8(3), 518–534. Larivière, S., Lavigne, K. M., Woodward, T. S., Gerretsen, P., Graff- Davis, K. D., Pope, G. E., Crawley, A. P., & Mikulis, D. J. (2002). Neural Guerrero, A., & Menon, M. (2017). Altered functional connectivity correlates of prickle sensation: A percept-related fMRI study. in brain networks underlying self-referential processing in delusions Nature Neuroscience, 5(11), 1121–1122. of reference in schizophrenia. Psychiatry Research: Neuroimaging, Diano, M., D’Agata, F., Cauda, F., Costa, T., Geda, E., Sacco, K., Duca, 263,32–43. S., Torta, D. M., & Geminiani, G. C. (2016). Cerebellar clustering Legrain, V., Iannetti, G. D., Plaghki, L., & Mouraux, A. (2011). The pain and functional connectivity during pain processing. Cerebellum, matrix reloaded: A salience detection system for the body. Progress 15(3), 343–356. in Neurobiology, 93(1), 111–124. Dixon, M. L., Andrews-Hanna, J. R., Spreng, R. N., Irving, Z. C., Mills, Lindquist, M. A. (2008). The statistical analysis of fMRI data. Statistical C., Girn, M., & Christoff, K. (2017). Interactions between the de- Science, 23(4), 439–464. fault network and dorsal attention network vary across default sub- Lindquist, M. A., Loh, J. M., Atlas, L. Y., & Wager, T. D. (2009). systems, time, and cognitive states. Neuroimage, 147,632–649. Modeling the hemodynamic response function in fMRI: Eccleston, C., Morley, S. J., & Williams, A. D. C. (2013). Psychological Efficiency, bias and mis-modeling. Neuroimage, 45(1), S187–S198. approaches to chronic pain management: Evidence and challenges. Loggia, M. L., Kim, J., Gollub, R. L., Vangel, M. G., Kirsch, I., Kong, J., British Journal of Anaesthesia, 111(1), 59–63. Wasan, A. D., & Napadow, V. (2013). Default mode network con- Fomberstein, K., Qadri, S., & Ramani, R. (2013). Functional MRI and nectivity encodes clinical pain: An arterial spin labeling study. pain. Current Opinion in Anesthesiology, 26(5), 588–593. PAIN®, 154(1), 24–33. Fox, J. (2015). Applied regression analysis and generalized linear May, A. (2008). Chronic pain may change the structure of the brain. models. Sage Publications. PAIN®, 137(1), 7–15. Goffaux, P., Girard-Tremblay, L., Marchand, S., Daigle, K., & Whittingstall, K. (2014). Individual differences in pain sensitivity Melzack, R., & Casey, K. L. (1968). Sensory, motivational, and central control determinants of pain: A new conceptual model. The Skin vary as a function of precuneus reactivity. Brain Topography, 27(3), 366–374. Senses, 1,423–443. 172 Neuroinform (2022) 20:155–172 Menon, V. (2015). Salience network. In A. W. Toga (Ed.), Brain map- Schweinhardt, P., & Bushnell, M. C. (2010). Pain imaging in health and disease—How far have we come? The Journal of Clinical ping: An encyclopedic reference (pp. 597–611). Academic Press: Elsevier. Investigation, 120(11), 3788–3797. Shulman, G. L., Fiez, J. A., Corbetta, M., Buckner, R. L., Miezin, F. M., Menon, V., & Uddin, L. Q. (2010). Saliency, switching, attention and Raichle, M. E., & Petersen, S. E. (1997). Common blood flow control: A network model of insula function. Brain Structure & changes across visual tasks: II. Decreases in cerebral cortex. Function, 214(5–6), 655–667. Journal of Cognitive Neuroscience, 9(5), 648–663. Metzak, P. D., Riley, J. D., Wang, L., Whitman, J. C., Ngan, E. T., & Sridharan, D., Levitin, D. J., & Menon, V. (2008). A critical role for the Woodward, T. S. (2011). Decreased efficiency of task-positive and right fronto-insular cortex in switching between central-executive task-negative networks during working memory in schizophrenia. and default-mode networks. Proceedings of the National Academy Schizophrenia Bulletin, 38(4), 803–813. of Sciences, 105(34), 12569–12574. Moayedi, M., Salomons, T. V., & Atlas, L. Y. (2018). Pain neuroimaging Takane, Y., & Hunter, M. A. (2001). Constrained principal component in humans: A primer for beginners and non-imagers. The Journal of analysis: A comprehensive theory. AAECC, 12(5), 391–419. Pain, 19(9), 961–9e1. Takane, Y., & Shibayama, T. (1991). Principal component analysis with Moulton, E. A., Pendse, G., Becerra, L. R., & Borsook, D. (2012). BOLD external information on both subjects and variables. Psychometrika, responses in somatosensory cortices better reflect heat sensation 56(1), 97–120. than pain. The Journal of Neuroscience, 32(17), 6024–6031. van der Miesen, M. M., Lindquist, M. A., & Wager, T. D. (2019). Mouraux, A., & Iannetti, G. D. (2018). The search for pain biomarkers in Neuroimaging-based biomarkers for pain: State of the field and the human brain. Brain, 141(12), 3290–3307. current directions. Pain Reports, 4(4). Peng, K., Steele, S. C., Becerra, L., & Borsook, D. (2018). Brodmann Vierck, C.J.,Whitsel,B.L.,Favorov,O.V.,Brown,A.W.,& area 10: collating, integrating and high level processing of Tommerdahl, M. (2013). Role of primary somatosensory cortex in nociception and pain. Progress in neurobiology, 161,1–22. the coding of pain. PAIN®, 154(3), 334–344. Percival, C.M., Zahid, H.B., & Woodward, T. S. (2020). CNoS-Lab/ Wager, T. D., Atlas, L. Y., Lindquist, M. A., Roy, M., Woo, C. W., & Woodward_Atlas. Zenodo. https://doi.org/10.5281/zenodo. Kross, E. (2013). An fMRI-based neurologic signature of physical pain. The New England Journal of Medicine, 368(15), 1388–1397. Peyron, R., García-Larrea, L., Grégoire, M. C., Costes, N., Convers, P., Weissman, D. H., Roberts, K. C., Visscher, K. M., & Woldorff, M. G. Lavenne, F., Mauguière, F., Michel, D., & Laurent, B. (1999). (2006). The neural bases of momentary lapses in attention. Nature Haemodynamic brain responses to acute pain in humans: Sensory Neuroscience, 9(7), 971–978. and attentional networks. Brain, 122(9), 1765–1780. Wilcox, C. E., Mayer, A. R., Teshiba, T. M., Ling, J., Smith, B. W., Peyron, R., Laurent, B., & Garcia-Larrea, L. (2000). Functional imaging Wilcox, G. L., & Mullins, P. G. (2015). The subjective experience of brain responses to pain. A review and meta-analysis (2000). of pain: An FMRI study of percept-related models and functional Neurophysiologie Clinique/Clinical Neurophysiology, 30(5), 263– connectivity. Pain Medicine, 16(11), 2121–2133. Woo, C. W., Roy, M., Buhle, J. T., & Wager, T. D. (2015). Distinct brain Postorino, M., May, E. S., Nickel, M. M., Tiemann, L., & Ploner, M. systems mediate the effects of nociceptive input and self-regulation (2017). Influence of pain on motor preparation in the human brain. on pain. PLoS Biology, 13(1), e1002036. Journal of Neurophysiology, 118(4), 2267–2274. Yam, M. F., Loh, Y. C., Tan, C. S., Khadijah Adam, S., Abdul Manan, N., & Basir, R. (2018). General pathways of pain sensation and the Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, major neurotransmitters involved in pain regulation. International D. A., & Shulman, G. L. (2001). A default mode of brain function. Journal of Molecular Sciences, 19(8), 2164. Proceedings of the National Academy of Sciences, 98(2), 676–682. Rosa, M. J., & Seymour, B. (2014). Decoding the matrix: Benefits and limitations of applying machine learning algorithms to pain neuro- Publisher’sNote Springer Nature remains neutral with regard to jurisdic- imaging. Pain, 155(5), 864–867. tional claims in published maps and institutional affiliations.

Journal

NeuroinformaticsSpringer Journals

Published: Jan 1, 2022

Keywords: Functional MRI; Functional brain networks; Functional connectivity; Pain; Multivariate least-squares regression; Principal component analysis; Hemodynamic responses; Attention

There are no references for this article.