Share this article on:

Multivariate machine learning distinguishes cross-network dynamic functional connectivity patterns in state and trait neuropathic pain

Cheng, Joshua C.a,b; Rogachov, Antona,b; Hemington, Kasey S.a,b; Kucyi, Aaronc; Bosma, Rachael L.a; Lindquist, Martin A.d; Inman, Robert D.b,e; Davis, Karen D.a,b,f,*

doi: 10.1097/j.pain.0000000000001264
Research Paper
Editor's Choice

Communication within the brain is dynamic. Chronic pain can also be dynamic, with varying intensities experienced over time. Little is known of how brain dynamics are disrupted in chronic pain, or relates to patients' pain assessed at various timescales (eg, short-term state vs long-term trait). Patients experience pain “traits” indicative of their general condition, but also pain “states” that vary day to day. Here, we used network-based multivariate machine learning to determine how patterns in dynamic and static brain communication are related to different characteristics and timescales of chronic pain. Our models were based on resting-state dynamic functional connectivity (dFC) and static functional connectivity in patients with chronic neuropathic pain (NP) or non-NP. The most prominent networks in the models were the default mode, salience, and executive control networks. We also found that cross-network measures of dFC rather than static functional connectivity were better associated with patients' pain, but only in those with NP features. These associations were also more highly and widely associated with measures of trait rather than state pain. Furthermore, greater dynamic connectivity with executive control networks was associated with milder NP, but greater dynamic connectivity with limbic networks was associated with greater NP. Compared with healthy individuals, the dFC features most highly related to trait NP were also more abnormal in patients with greater pain. Our findings indicate that dFC reflects patients' overall pain condition (ie, trait pain), not just their current state, and is impacted by complexities in pain features beyond intensity.

Using machine learning, we identified abnormal dynamic functional connectivity across multiple brain networks that were linked to neuropathic trait pain.

aDivision of Brain, Imaging, and Behaviour—Systems Neuroscience, Krembil Research Institute, Toronto Western Hospital, University Health Network, Toronto, ON, Canada

bInstitute of Medical Science, University of Toronto, Toronto, ON, Canada

cDepartment of Neurology and Neurological Sciences, Stanford University, Stanford, CA, United States

dDepartment of Biostatistics, Johns Hopkins University, United States

Departments of eMedicine and

fSurgery, University of Toronto, Toronto, ON, Canada

Corresponding author. Address: Krembil Research Institute, Toronto Western Hospital, 399 Bathurst St, Room MP12-306, Toronto, ON M5T 2S8, Canada. Tel.: (416) 603-5662. E-mail address: karen.davis@uhnresearch.ca (K.D. Davis).

Sponsorships or competing interests that may be relevant to content are disclosed at the end of this article.

Received February 16, 2018

Received in revised form April 09, 2018

Accepted April 23, 2018

Back to Top | Article Outline

1. Introduction

Chronic pain, which is the persistence of pain beyond 3 to 6 months, is estimated to affect 20% of the population.36 Chronic pain can fluctuate in intensity and character over time. Thus, a current pain rating reflects a state of pain in the moment (“state pain”), but may not capture the overall intensity/chronicity of the pain in general, which is better described by an average pain rating over a longer period (“trait pain”). It is not known whether brain measures obtained during neuroimaging better reflect state or trait symptom assessments of pain. Understanding brain representations of state and trait pains provide fundamental insight into chronic pain.

In the brain, the representation of chronic pain is complex and dynamic.21,22 It not only involves the regional activity within a multitude of brain areas, but the dynamic coordination of this activity within and across networks.10,22 Previous functional magnetic resonance imaging (MRI) studies identified alterations in properties of brain communication in patients with chronic pain, as represented by “static” functional connectivity (sFC) measures.1,16,25,30 However, measurements of moment-by-moment synchrony on a finer timescale can capture the dynamic interplay between brain regions.17 One such measure is dynamic functional connectivity (dFC), which measures fluctuations in correlated activity over time. Currently, it is not well understood how measures of dFC within and/or across networks relate to chronic pain.

In addition, patients within cohorts ascribed the same clinical diagnosis may exhibit variability in their underlying pain pathophysiology. For example, we examined patients with chronic pain suffering from ankylosing spondylitis (AS)—an inflammatory arthritis that affects the axial skeleton.32 Ankylosing spondylitis–related pain has traditionally been viewed to be inflammatory mediated,3 but a subset of patients have neuropathic pain (NP).40 Given the differences in pathophysiology that exists between inflammatory pain and NP,41 it is important to consider whether the relationship between brain activity and AS-related pain is pathophysiology dependent. For example, NP has a temporal dimension in which patients experience fluctuating, spontaneous, and ongoing pain. Therefore, measures of brain dynamics may be more sensitive to capture these temporal features of pain, vs others.

Here, using a multivariate machine learning approach, our goals were to (1) determine whether multivariate dFC and sFC patterns better reflects state or trait measures of chronic pain, and (2) determine whether these brain patterns are influenced by complexities in pain features beyond intensity (ie, neuropathic vs non-neuropathic). Based on previous findings of abnormal cross-network connectivity in patients with AS (between the default mode network [DMN] and salience network [SN]),16 we hypothesized that (1) AS-related pain is related to cross-network but not within-network brain patterns of dFC and sFC. We also hypothesized that (2) the multivariate brain models for state vs trait AS pain would differ, (3) the sensitivity of these models will be driven by patients with NP features, and (4) brain measures best associated with AS pain will demonstrate the degree of departure from healthy brain function, such that patients with greater pain will have greater brain abnormalities.

Back to Top | Article Outline

2. Methods

2.1. Subjects

We collected psychophysical and neuroimaging data from 71 patients with AS (51 men and 20 women, ages 18-61) and 62 age-/sex-matched healthy controls (43 men and 19 women, ages 18-55). A subset of these data was used in a different study for another experimental question.15 All subjects provided informed written consent to procedures approved by the University Health Network Research Ethics Board. Inclusion criteria for the patients included: (1) a diagnosis of AS according to the 1984 Modified New York Criteria,37 (2) received only nonsteroidal anti-inflammatory drugs or biologic agents for treatment of AS, and (3) experiencing AS-related pain. Exclusion criteria for both patients and healthy controls were (1) any contraindication for having an MRI, (2) a history of psychiatric, neurologic, or metabolic conditions, or (3) major surgery within the past 2 years.

Back to Top | Article Outline

2.2. Clinical measures: pain ratings, painDETECT

All but one patient completed the painDETECT questionnaire after their scan11—a screening form designed to detect NP. The total painDETECT score was calculated according to the specified instructions.11 The painDETECT was also used to assess current pain as a proxy for “state pain,” and average monthly pain as a proxy for “trait pain.” To do this, we documented the responses to the questions: (1) “How would you assess your pain now, at this moment?” (0-10; none-max), and (2) “How strong was the pain during the past 4 weeks on average?” (0-10; none-max). A total of 9 subjects rated their current pain to be 0, and 3 subjects did not answer this question, and so, only 58 subjects were used in the analyses of state pain. A total of 4 subjects rated their 4-week pain to be 0, and the same 3 subjects who did not rate their current pain also did not rate their 4-week pain. Thus, a total of 63 subjects were included in analyses of trait pain.

Back to Top | Article Outline

2.3. Classification of neuropathic pain

The painDETECT questionnaire is constructed to distinguish the likelihood that a patient has NP, with scores ranging from 0 to 39; scores ≤12 indicating that NP is unlikely to be present and scores ≥19 indicating that an NP component is present.11 painDETECT scores between 13 and 18 indicate that a neuropathic component may be present.11 Thus, as per done previously,40 we categorized patients with AS into a low painDETECT group (scores ≤12; non-neuropathic [non-NP]) and a high painDETECT group (scores >12; neuropathic [NP]).

Back to Top | Article Outline

2.4. Neuroimaging acquisition

Each subject underwent an MRI (3T GE) brain imaging session, with the scanner fitted with an 8-channel phased-array head coil. From each subject, we acquired: (1) a T1-weighted anatomical scan (3D IR-FSPGR sequence, 1 × 1 × 1 mm3 voxels, matrix = 256 × 256, 180 axial slices, repetition time = 7.8 ms, echo time = 3 ms, and inversion time = 450 ms), and (2) a 9-minute 14-second T2*-weighted resting-state fMRI scan (echo-planar imaging sequence, 3.125 × 3.125 × 4 mm3 voxels, matrix = 64 × 64, 36 axial slices, repetition time = 2 seconds, echo time = 30 ms, flip angle = 85°, 277 volumes, interleaved slice acquisition, no gap, and slice thickness = 4 mm). The instructions provided for the resting-state scan was “close your eyes, do not try to think about anything in particular; do not fall asleep.”

Back to Top | Article Outline

2.5. Resting-state functional magnetic resonance imaging preprocessing

The first 4 volumes of the resting-state scan were removed using FEAT from FSL v5.0.18 Nonbrain voxels were removed using the Brain Extract Tool (BET) within FEAT, and motion correction was performed using MCFLIRT. Each subject's resting-state image was then linearly registered to their T1-weighted anatomical image, which was skull stripped using the optiBET tool,28 followed by nonlinear registration to MNI152-2mm space using FNIRT. aCompCor2,5 was then performed as per previously reported,24 to remove scanner-related and physiological noise. The top 5 white matter and cerebrospinal fluid components were regressed out, as well as the 6 motion parameters. Spatial smoothing was performed using a 6-mm full-width at half-maximum kernel, followed by temporal filtering in FSL to retain the signal between 0.01 and 0.1 Hz.

Back to Top | Article Outline

2.6. Individualized parcellation

Each subject's preprocessed resting-state data were parcellated into 18 individualized resting-state networks (Fig. 1) separately for their right and left hemisphere, using a published parcellation pipeline (http://nmr.mgh.harvard.edu/bid/DownLoad.html), which has previously been described in detail.39 Briefly, each subject's T1-weighted anatomical image was preprocessed in FreeSurfer v6 using the recon-all function. Each subject's preprocessed resting-state fMRI image was then registered to their structural image, followed by alignment to a common FreeSurfer template (fsaverage6; 40,962 vertices in each hemisphere). Smoothing with a 6-mm full-width at half-maximum kernel was then applied in surface space, followed by downsampling of these data to fsaverage4 space (2562 vertices in each hemisphere). For the individualized parcellation procedure, an 18-network group-level atlas was projected onto each subject's cortical surface, and the average time series for each network was then taken for each individual—termed the reference signal. Thus, each individual had 18 reference resting-state fMRI signals corresponding to each of their 18 networks. Subjects' resting-state time series at each vertex was then correlated with each of their 18 reference signals and assigned to the network with the maximum correlation. For each vertex, the ratio between the largest and second largest correlation values was then taken as a measure of confidence. In addition, several parameters including signal-to-noise and preestimated intersubject variability in FC were also taken into account.39 The vertices that had a confidence value >1.1 were then averaged for each network, termed the core signal, and multiplied with the parameters described previously. This signal was then averaged with the original reference signal in a weighted manner and served as the reference signal for the next iteration. This iterative procedure was halted when 98% of the vertices retained the same membership after 2 consecutive iterations.

Figure 1

Figure 1

The first 17 networks correspond to a previously published parcellation,43 with the 18th network corresponding to the hand sensorimotor network.39 Although the original 17-network parcellation was not named, a previous publication33 overlaid the 17 networks onto the 7-network Yeo parcellation which was named.43 We followed this naming convention with the exception of network 4—which better corresponded to the dorsal posterior insula (dpIns) and parietal operculum rather than somatomotor network—and network 8—which could not be distinguished between salience and executive control.

Back to Top | Article Outline

2.7. Group-level networks

For visualization purposes, group-level networks (Fig. 2 and Table 1) were also derived based on all the subjects who were included in the state or trait pain analyses (64 subjects total). Each subject's functional parcellation was registered into volumetric space, followed by nonlinear registration to standard MNI152-2mm space. For each network, subject maps were binarized and added together using fslmaths. These group-level maps were then thresholded at 80% (∼51/64 subjects) to visualize areas where a majority of the subjects overlapped in their individually derived networks.

Figure 2

Figure 2

Table 1

Table 1

Back to Top | Article Outline

2.8. Within-network static functional connectivity and dynamic functional connectivity

The individualized network parcellation39 outputs network rather than node-level labels. Thus, to calculate within-network connectivity measures for each subject, each network was first split into constrained anatomical nodes. This was performed by binarizing each network label in Freesurfer for each individual, followed by the mri_surfcluster command. Consequently, the number of nodes for each network could differ between patients. To calculate within-network sFC, average resting-state time series were extracted from nodes of each network across both hemispheres. For each of the 18 networks in each patient, pairwise Pearson correlations were performed between time series of every pair of nodes within the network, followed by Fisher r-to-z transformation of these correlation values. These measures were then averaged for each network and for each subject. Thus, each subject had 18 within-network sFC values corresponding to each of their 18 individualized networks. For within-network dFC, the dynamic conditional correlation method was applied between time series of every pair of nodes within each network. The SD of each dynamic conditional correlation was then taken as a measure of dFC as per done previously6 and shown to be test–retest reliable.8 These dFC values were then averaged within each network. Thus, each subject had 18 within-network dFC values corresponding to each of their 18 individualized networks.

Back to Top | Article Outline

2.9. Cross-network static functional connectivity and dynamic functional connectivity

The average time series for each of the 18 individualized resting-state networks across the brain was extracted for each subject. To calculate cross-network sFC in each subject, Pearson correlations were performed between averaged time series of every pair of networks, followed by Fisher r-to-z transformation of these correlation values. Thus, 153 unique internetwork sFC measures were generated for each subject. To calculate internetwork dFC in each subject, the dynamic conditional correlation method was applied between averaged time series of every pair of networks. The SD of each dynamic conditional correlation was then taken as a measure of dFC. Thus, 153 unique internetwork dynamic connectivity measures were generated for each subject.

Back to Top | Article Outline

2.10. Machine learning: elastic-net regularized cross-validated regression

Multiple linear regression is a useful technique for analyzing the association between many independent variables and a continuous dependent variable. However, given the presence of many independent variables (dFC and sFC in this study), overfitting and problems with multicollinearity can arise.19 One solution is to include only the relevant independent variables in the model by penalizing greater model complexity—a technique called regularization.42 Thus, we used elastic-net regression45—a type of regularized multiple linear regression—to look at the association between brain connectivity (dFC and sFC) and pain (state and trait) (Fig. 3 for methods schematic). To perform the necessary parameter tuning involved with elastic-net regression, and to test the generalizability of the outputted multivariate brain model for pain, we implemented elastic-net regression within a nested cross-validation framework. In a nested cross-validation, the whole data set is split into a training set and a testing set for each outer fold. The training set in the outer fold is then further split into a training set and a testing set for each inner fold (Fig. 3). During inner fold cross-validation, optimization of parameters for elastic-net (alpha and lambda) is performed, and the multivariate model of brain connectivity for pain is created. The value of lambda determines the amount of regularization that is performed, and a range of 50 values to cross-validate across from ∼0.01 to 1 was supplied. Alpha determines whether the model is lasso-regularized (alpha = 1; regression coefficients can be set to 0), ridge-regularized (alpha = 0; regression coefficients are minimized but not equal to 0), or in-between (elastic-net). The 3 values of alpha cross-validated across were 0.1, 0.5, and 1. A value 0.1 rather than 0 was chosen as a lower limit of alpha, as we did not expect every independent variable to be associated with the outcome variable, and thus wanted at least a minimal level of feature selection. Thus, a total possible 150 parameter combinations were tested within the inner loop. The combination of alpha and lambda that led to the lowest average mean squared error across folds was determined within the inner loop, and then the entire training set of the outer loop was trained using this parameter combination to derive the multivariate brain model of pain. To determine the goodness of this model, or the ability to generalize to new data, this model was tested on the testing data within the outer fold. Here, the brain connectivity features of the testing subjects were provided, but their self-reported pain ratings were not. If the multivariate model developed within the inner fold captured generalizable associations between brain connectivity features and pain, the model estimates of pain based on the testing subjects' brain connectivity would be positively correlated with their self-reported pain ratings. Within both the inner and the outer folds, K-fold rather than leave-one-out cross-validation was used, as leave-one-out cross-validation suffers from higher variance and thus does not yield stable estimates of predictive accuracy.38 For analyses that used all subjects, 20-fold cross-validation was used within the outer loop, and 10-fold cross-validation was used within the inner loop. With analyses that looked separately at patients with non-NP and NP, 10-fold cross-validation was used within the outer loop because of a fewer number of subjects, and 10-fold cross-validation was used within the inner loop.

Figure 3

Figure 3

As a measure of performance, a Spearman correlation was run between model estimates of the outcome variable across all folds of the outer loop, with actual values of the outcome variable itself. Statistical significance testing of the models was performed using permutation-based testing, with significance set at P < 0.05. Permutation-based significance testing was only performed in models where a positive correlation was observed between model estimates and actual ratings. If a negative correlation was observed between model estimates and actual ratings, then these models were not statistically significant (P > 0.05). For permutation-based significance testing, the dependent variable (self-reported pain) was shuffled from the independent variables (dFC and sFC) across subjects through random permutation, and the entire machine learning procedure was repeated. This was iterated 2500 times. The P value was calculated by finding the proportion of iterations in which the shuffled model had a larger correlation value between model estimates and actual ratings, in comparison with the correlation value derived from the unpermuted model.

We derived 12 multivariate brain connectivity models for pain in total. Two examined the relationship between cross-network brain measures with state and trait pain, and 2 examined the relationship between within-network brain measures with state and trait pain. Each analysis was then repeated for patients with NP and non-NP separately to see whether the model differed according to these groups.

Back to Top | Article Outline

2.11. Feature extraction and correlation with forward model

The magnitude of the multivariate weight for each feature can represent the importance of the feature for the model. Here, we extracted the multivariate weights to determine the features that drove the significant brain models of AS-related pain. To do this, the median regression coefficient for each feature was taken across all the outer-fold models. This was done because with each elastic-net regression run, the number of models generated was equivalent to the number of outer folds. However, it has also been shown that the magnitude of a multivariate weight may not be related to the amount of “signal” (relationship with state pain here) for each feature.14 Rather, a forward modeling approach, such as examining the covariance between each feature (dFC and sFC) and the dependent variable (pain), may be more appropriate.12,14 Thus, we also compared the multivariate weights with the covariance pattern before we interpreted the magnitude of the multivariate weights as a proxy of feature importance. To do so, we compared the median multivariate weight for each feature with the covariance of that feature with pain to determine whether their sign (positive/negative) was in the same direction. We also ran a Spearman correlation between the multivariate weight pattern with the covariance pattern to determine whether they were similar or different.

Back to Top | Article Outline

2.12. Comparison of brain features between patients with neuropathic pain and healthy controls

As the cross-network brain models for state and trait NP were statistically significant (see Results), we investigated whether the magnitude of NP resulted in more prominent abnormalities across the top-10 features of these models compared with healthy controls. To do so, we compared patterns of brain connectivity across these top features between patients with NP and their age-/sex-matched healthy controls. For the top-10 features selected in a cross-network brain model for pain, the appropriate dFC and sFC values were extracted for each patient with NP and their age-/sex-matched healthy control. Separately for each patient and their matched healthy control, a Spearman correlation was then run between these extracted features. This yielded a correlation value for each patient, representing the similarity across these features to their matched healthy control. A Spearman correlation was then used between this similarity measure and patients' pain (state or trait depending on the model assessed) across all patients with NP. This analysis was repeated for both the cross-network brain model for state and trait pain.

Back to Top | Article Outline

3. Results

3.1. Dissociation between pain intensity and pain pathophysiology

Across all patients, state pain ranged from 1 to 9, with a mean ± SD of 3.3 ± 2.2. Trait pain ratings were slightly higher on average across subjects, ranging from 1 to 8, with a mean ± SD of 4.0 ± 2.2. State and trait pain ratings were significantly positively correlated across all patients (Spearman rho = 0.74, P = 6.36 × 10−11). Large interindividual differences were observed in painDETECT ratings, ranging from 0 to 24, with a mean ± SD of 9.4 ± 5.5. The NP and non-NP groups were not statistically different in terms of state pain (meanNP ± SDNP = 4.1 ± 2.4, n = 19; meannon-NP ± SDnon-NP = 2.9 ± 2.0, n = 39; t = 1.9, P = 0.074) or trait pain (meanNP ± SDNP = 4.6 ± 2.0, n = 20; meannon-NP ± SDnon-NP = 3.8 ± 2.3, n = 43; t = 1.4, P = 0.18). The patients with NP who provided state pain ratings included 14 men and 5 women with a mean ± SD age of 38.9 ± 9.8, whereas the patients with NP who provided trait pain ratings included the previous subjects with an additional woman and a mean ± SD age of 39.0 ± 9.5. The patients with non-NP who provided state pain ratings included 29 men and 10 women with a mean ± SD age of 33.1 ± 11.3, whereas the patients with non-NP who provided trait pain ratings included 30 men and 13 women with a mean ± SD age of 32.1 ± 11.2.

Furthermore, there were no sex differences in state pain (meanfemales ± SDfemales = 4.2 ± 2.6, n = 15; meanmales ± SDmales = 3.1 ± 2.0, n = 43; t = 1.5, P = 0.14), trait pain (meanfemales ± SDfemales = 4.3 ± 2.4, n = 19; meanmales ± SDmales = 3.9 ± 2.2, n = 44; t = −0.62, P = 0.54), or painDETECT scores (meanfemales ± SDfemales = 9.9 ± 4.8, n = 20; meanmales ± SDmales = 9.1 ± 5.8, n = 50; t = −0.57, P = 0.58).

Back to Top | Article Outline

3.2. State pain can be modeled by multivariate cross-network but not within-network functional connectivity

We found a profound difference in the generalizability of models of state pain based on within-network connectivity vs cross-network connectivity. The multivariate brain model for state pain derived using within-network sFC and dFC failed to generalize to unseen (left-out) subjects (Spearman rho = 0.21, P = 0.17). By contrast, the multivariate brain model for state pain derived using cross-network sFC and dFC generalized to unseen subjects, with model-estimated state pain being significantly correlated with self-reported state pain (model 1 in Fig. 4; Spearman rho = 0.41, P = 0.01). In Figure 4, we further show how the model estimate of pain differs from the self-report of pain for each individual. The deviation of the regression line from a perfect model (ie, the dotted line where y = x) indicates model overestimates and underestimates of pain. For state pain, these 2 lines cross at a level of approximately 3/10 pain and the y-intercept is approximately 2.5. Thus, for subjects who reported a state pain ∼ <3, the model tended to overestimate their pain, whereas the model tended to underestimate pain for subjects who reported a state pain >3 (model 1 in Fig. 4).

Figure 4

Figure 4

To determine whether the generalizability of the cross-network model of state pain was driven by a particular type of pain, we constructed separate cross-network models for the patients with NP (model 2) and non-NP (model 3) (Fig. 4). These specific pain type models indicate that model 1 was likely driven by the patients with NP. Specifically, there was a stronger relationship between model-estimated state pain and self-reported state pain within NP patients alone (model 2 in Fig. 4; Spearman rho = 0.55, P = 0.03) than within non-NP patients alone (model 3 in Fig. 4; Spearman rho = −0.09, P > 0.05), and all subjects combined (model 1: Spearman rho = 0.41). By contrast, the ability of model 1 to generalize to unseen subjects was not driven by sex (women: Spearman rho = −0.27, P > 0.05; men: Spearman rho = 0.05, P = 0.16) or treatment effects (no biologics: Spearman rho = 0.27, P = 0.12; biologics: Spearman rho = −0.06, P > 0.05).

In examining the pain estimates in model 2, it was similar to model 1 in which it overestimated low levels of self-reported pain and underestimated higher levels of self-reported pain. For completion, we also constructed within-network models of state pain in NP and non-NP patients alone, but these models did not generalize to unseen subjects (Spearman rho = −0.11, P > 0.05; Spearman rho = −0.18, P > 0.05, respectively).

Back to Top | Article Outline

3.3. Features of cross-network functional connectivity that drive the multivariate model of state neuropathic pain

We found that the multivariate weight for each feature selected always had the same sign (positive/negative) as the covariance of that feature with state pain. Furthermore, the multivariate weights were highly correlated with the covariances across the selected features (Spearman rho = 0.74, P = 2.23 × 10−7, n = 41). In comparing the cross-network dFC and sFC features (Fig. 5), cross-network dFC features generally had larger magnitude weights than the sFC features selected (absolute meandFC ± SDdFC = 0.85 ± 0.87, n = 21, absolute meansFC ± SDsFC = 0.34 ± 0.35, n = 20, t = 2.5, P = 0.019). Examining these weight patterns, greater cross-network dFC was also typically associated with less pain (negative regression weights), whereas greater sFC was typically associated with more pain (positive regression weights). For each feature selected, we also plotted the dFC and sFC for a single patient with NP, and for comparison, the dFC and sFC for these features in an age-/sex-matched healthy control (Fig. 5).

Figure 5

Figure 5

Back to Top | Article Outline

3.4. Multivariate cross-network functional connectivity models trait pain in patients with neuropathic pain but not non–neuropathic pain

When we investigated whether cross- or within-network brain connectivity was associated with patients' trait pain, we found that the multivariate cross-network (model 4 in Fig. 4) and within-network brain models for trait pain derived across all subjects did not generalize to unseen subjects (Spearman rho = −0.22, P > 0.05; Spearman rho = −0.28, P > 0.05, respectively). However, the multivariate cross-network brain model for trait pain derived in patients with NP (model 5 in Fig. 4) but not non-NP (model 6 in Fig. 4) was able to generalize to unseen subjects (Spearman rho = 0.72, P = 0.004; Spearman rho = −0.08, P > 0.05, respectively). In examining the pain estimates in model 5, it was similar to model 2 in which it overestimated low levels of self-reported pain and underestimated higher levels of self-reported pain.

We next examined the features selected in the cross-network brain model for trait pain derived in patients with NP (model 5). The multivariate weight for each feature always had the same sign as the covariance of that feature with trait pain. The multivariate weights were also strongly correlated with the covariances across the selected features (Spearman rho = 0.48, P = 1.25 × 10−6, n = 93). Similar to model 2, cross-network dFC features generally had larger magnitude weights than the sFC features selected (absolute meandFC ± SDdFC = 0.80 ± 0.75, n = 43, absolute meansFC ± SDsFC = 0.14 ± 0.17, n = 50, t = 5.64, P = 1.0 × 10−6) (Fig. 6). However, different from model 2 was that more features were selected in the trait pain compared with the state pain model (n = 93 vs 41; Fig. 6 vs Fig. 5, respectively), reflected by a more ridge-like regularization penalty selected in the model (median alpha = 0.1 vs 0.5, respectively).

Figure 6

Figure 6

For completion, we also generated multivariate within-network brain models for trait pain in patients with NP and patients with non-NP, but these models were unable to generalize to unseen subjects in their respective groups (Spearman rho = −0.40, P > 0.05; Spearman rho = −0.13, P > 0.05, respectively).

Back to Top | Article Outline

3.5. Cross-network brain models for state and trait neuropathic pain are not driven by head motion

As head motion may cause spurious correlations in FC,31 we also repeated the model building of models 2 and 5 but with mean relative head displacement for each subject included as an additional independent variable. Mean relative head displacement was not selected as a feature in either of these models. In addition, similar correlations were observed between model estimates and self-report for the state and trait NP models when mean relative head displacement was accounted for (Spearman rho = 0.52, P = 0.035 and Spearman rho = 0.78, P = 0.0016, respectively), compared with when it was not (Spearman rho = 0.55, P = 0.03 and Spearman rho = 0.72, P = 0.004, respectively).

Back to Top | Article Outline

3.6. Top features in the multivariate cross-network brain models for neuropathic pain are dynamic functional connectivity features

To visualize the top-10 most important features in the cross-network brain models for state and trait NP (models 2 and 5, respectively), we created circle plots (Fig. 7; Figs. 1 and 2 for corresponding network visualization). For the cross-network brain model of state pain, 8/10 top features were cross-network dFC and 2 were cross-network sFC (Table 2 and Fig. 7). Of these top-10 dFC features, cross-network dFC with executive control networks (ECN with visual, somatomotor, limbic networks as well as the DMN, and dpINS) was negatively related to state pain, whereas cross-network dFC with limbic networks (limbic with attention and SN/ECN) was positively related to state pain. Of the sFC features (DMN with the attention and somatomotor networks) within the top-10, cross-network sFC was positively related to state pain.

Figure 7

Figure 7

Table 2

Table 2

For the cross-network brain model of trait pain, all the features with the top-10 greatest multivariate weights were dFC features (Table 2 and Fig. 7). Similar to findings with state pain, cross-network dFC with ECNs (ECN with visual, somatomotor, attention networks as well as the dpINS) was negatively related to trait pain, and cross-network dFC with limbic networks (limbic with attention and SN/ECN) was positively related to trait pain. Unique to trait pain, however, was that cross-network dFC with DMNs (DMN with somatomotor and SN/ECN as well as the dpINS) was additionally positively related to trait pain.

Back to Top | Article Outline

3.7. Greater magnitude of trait pain in patients with neuropathic pain is associated with greater brain abnormalities

We also investigated whether the magnitude of NP resulted in more prominent abnormalities across the top-10 features of the significant brain models for state and trait pain compared with age-/sex-matched healthy controls. There were no statistically significant differences in age between patients with NP and matched healthy controls for the state pain analysis (meanNP ± SDNP = 38.9 ± 9.8; meancontrols ± SDcontrols = 37.8 ± 11.1; P = 0.75, t = 0.32) nor for the trait pain analysis (meanNP ± SDNP = 39.0 ± 9.5; meancontrols ± SDcontrols = 37.9 ± 10.8; P = 0.74, t = 0.34). We found that for the top-10 features selected in the cross-network brain model for trait pain (model 5), NP patients with milder trait pain showed more correspondence with their matched healthy control than patients with greater pain (Fig. 7; Spearman rho = −0.58, P = 0.008). Although a similar trend was observed with the top-10 features selected in the brain model for state pain (model 2), this relationship was not statistically significant (Spearman rho = −0.38, P = 0.11).

Back to Top | Article Outline

4. Discussion

This is the first study to reveal how dynamics in brain communication relate to fundamental characteristics and timescales of chronic pain, using a machine learning approach. Our key findings were (1) the most prominent networks in our models were the default mode, salience, and ECNs, (2) state and trait chronic pain are related to patterns of cross-network but not within-network dFC and sFC, (3) the cross-network brain–pain relationships were present in patients with NP but not non-NP, (4) dynamic connectivity patterns form the core of multivariate cross-network models of chronic pain in AS with only a minor contribution from sFC features, (5) the cross-network brain model for trait pain is more complex and comprised of more features than the model for state pain, and (6) patients with NP with milder trait pain show greater brain network similarity to healthy controls than patients with greater pain. This study demonstrates that measures of brain dynamics most sensitively reflects trait pain in patients with NP features.

We and others have previously shown that patients with chronic pain can exhibit altered cross-network sFC.1,16,30 However, brain communication and FC fluctuates over various timescales,17,20 and the characterization of dynamics provides information not captured by sFC.26 This is especially pertinent in the scope of cross-network communication, as intermodular/network connections are typically the most dynamic across the brain.44 Thus, this may underlie the finding here that most of the top features within the cross-network brain model of pain were dFC features. By contrast, patterns of within-network dFC and sFC were not associated with chronic state or trait pain. These findings may seem contradictory to other previous studies, which have found abnormalities in within-network functional communication in patients with chronic pain other than AS.4,23 However, we used an individualized functional mapping approach to delineate networks separately for each patient. Thus, abnormalities in within-network functional communication related to patients' pain may have been accounted for during the mapping of these individualized networks.

Also, previously unknown was whether these brain abnormalities represent a general state of chronic pain (ie, trait pain) or the patients' level of pain during the day of the scan (ie, state pain), which often fluctuates. In comparing the cross-network brain models for state and trait pain in patients with NP, key differences were found. The first was that many more brain connectivity features were selected in the multivariate model for trait rather than state pain. Second, the correlation between self-reported pain and model estimates made by the cross-network brain model of trait pain was marginally higher than that for state pain (Spearman rho = 0.72 vs 0.55, respectively). Thus, measures of patients' brain communication at rest were more highly and widely related to their trait compared with state pain. In comparison with healthy individuals, patients' brain networks related to trait rather than state pain were also increasingly abnormal with the level of their pain. One hypothesis for this amalgamation of findings is that as patients' pain often fluctuates over time, an assessment of pain averaged over a longer period may be more indicative of their current condition than an assessment of pain of the moment. Measures of resting-state dFC and sFC taken across the entire scan have also been shown to be test–retest reliable across neuroimaging sessions within individuals,7,8 capturing trait-like properties. Accordingly, trait-like brain measures captured during the resting-state was better able to reflect trait-like pain features across patients.

When the top features in the cross-network brain models for state and trait pain were examined, greater dFC with ECN was associated with less pain. This is consistent with our previous finding in healthy individuals, where greater cross-network dFC between the ECN and SN was associated with a better ability to cope with pain—as proxied by the individual's ability to prioritize cognitive task performance over attention to pain.6 In chronic pain, the dorsolateral prefrontal cortex, a hub of the ECN,34 has been shown to exhibit decreased gray matter across a broad range of chronic pain conditions.35 This region, typically ascribed to play an important role in cognitive control,29 also plays an important role in pain suppression.35 Thus, greater dynamic engagement of this area and its broader network in patients may be associated with a better ability to suppress pain. By contrast, greater dFC with limbic networks was associated with greater pain. This finding is consistent with a previous study on patients with persistent chronic back pain, which demonstrated that the brain representation of chronic pain predominated in emotion-/reward-related circuitry.13

Here, we found that although a multivariate model for state pain could be generated across all subjects, this was driven by patients with NP. Likewise, trait pain could be modelled by cross-network brain features in patients with NP only, and not in patients with non-NP or all patients together. This finding may reflect a greater sensitivity of dynamic measures of brain communication to behavioral measures that encompass dynamics as well, as NP typically involves fluctuating, spontaneous, and ongoing pain. Perhaps, in addition to sFC and dFC measures, other brain measures such as inflammation-linked proteins detected by PET,27 may be more sensitive to determine brain models that are highly associated with inflammatory pain or non-NP. Altogether, these findings highlight the importance of considering additional factors such as the type of pain presented when developing multivariate brain models for chronic pain, as heterogeneity in pain types may exist across patients grouped within the same diagnostic criteria. However, a limitation of this study was that we did not determine how well the findings would generalize for data that are acquired from a different scanner, or with different acquisition parameters.

Finally but importantly is the issue of the application of multivariate brain models of chronic pain for the diagnosis of chronic pain and treatment prognostication. The legal and ethical ramifications surrounding this topic have recently been discussed, and recommendations for the use of functional MRI offered by an international task force of experts.9 Findings from our study here support the task force findings that the application of machine learning and multivariate modeling to neuroimaging can provide great insight into the pathophysiology of chronic pain but should not be used to infer the presence or level of pain in an individual patient. Specifically, we found that the assessment of multivariate brain models on unseen/testing subjects in cross-validation was important for evaluating the goodness of the model, before interpreting associations between the independent (eg, brain measures) and dependent (pain) variables in the model. Although permutation testing showed that the cross-network brain models of AS-related pain in patients with NP were statistically significant in this study, predictions were not consistently accurate. Specifically, there was a tendency of these models to overestimate pain in patients with mild pain, and underestimate pain in patients with moderate to high levels of pain. Thus, although these models are useful for examining the association between multivariate patterns in brain connectivity and pain, estimates of pain in unseen subjects used for model evaluation purposes should not currently be used in lieu of self-reported pain in the clinic—which remains the gold standard.9

Back to Top | Article Outline

Conflict of interest statement

The authors have no conflict of interest to declare.

This work was supported by the Canadian Institute of Health Research (operating grant to K.D.D.); Strategy for Patient-Oriented Research (SPOR) funding of the Canadian Chronic Pain Network; and The Mayday Fund. J. Cheng and K. Hemington are recipients of a Canadian Institute of Health Research Doctoral Research Award. A. Kucyi was supported by a Banting fellowship from Canadian Institute of Health Research.

Back to Top | Article Outline

Acknowledgements

The authors thank Dr Adrian Crawley, Dr Judith Hunter, and Dr Paul Dufort for helpful insights and technical assistance with data analysis and interpretation, and Eugen Hlasny and Keith Ta for expert technical assistance in MRI acquisition.

Back to Top | Article Outline

References

[1]. Baliki MN, Geha PY, Apkarian AV, Chialvo DR. Beyond feeling: chronic pain hurts the brain, disrupting the default-mode network dynamics. J Neurosci 2008;28:1398–403.
[2]. Behzadi Y, Restom K, Liau J, Liu TT. A component based noise correction method (CompCor) for BOLD and perfusion based fMRI. Neuroimage 2007;37:90–101.
[3]. Bidad K, Gracey E, Hemington KS, Mapplebeck JCS, Davis KD, Inman RD. Pain in ankylosing spondylitis: a neuro-immune collaboration. Nat Rev Rheumatol 2017;13:410–20.
[4]. Bolwerk A, Seifert F, Maihofner C. Altered resting-state functional connectivity in complex regional pain syndrome. J Pain 2013;14:1107–15.e1108.
[5]. Chai XJ, Castanon AN, Ongur D, Whitfield-Gabrieli S. Anticorrelations in resting state networks without global signal regression. Neuroimage 2012;59:1420–8.
[6]. Cheng JC, Bosma RL, Hemington KS, Kucyi A, Lindquist MA, Davis KD. Slow-5 dynamic functional connectivity reflects the capacity to sustain cognitive performance during pain. Neuroimage 2017;157:61–8.
[7]. Choe AS, Jones CK, Joel SE, Muschelli J, Belegu V, Caffo BS, Lindquist MA, van Zijl PC, Pekar JJ. Reproducibility, temporal structure in weekly resting-state fMRI over a period of 3.5 years. PLoS One 2015;10:e0140134.
[8]. Choe AS, Nebel MB, Barber AD, Cohen JR, Xu Y, Pekar JJ, Caffo B, Lindquist MA. Comparing test-retest reliability of dynamic functional connectivity methods. Neuroimage 2017;158:155–75.
[9]. Davis KD, Flor H, Greely HT, Iannetti GD, Mackey S, Ploner M, Pustilnik A, Tracey I, Treede RD, Wager TD. Brain imaging tests for chronic pain: medical, legal and ethical issues and recommendations. Nat Rev Neurol 2017;13:624–38.
[10]. Davis KD, Moayedi M. Central mechanisms of pain revealed through functional and structural MRI. J Neuroimmune Pharmacol 2013;8:518–34.
[11]. Freynhagen R, Baron R, Gockel U, Tolle TR. painDETECT: a new screening questionnaire to identify neuropathic components in patients with back pain. Curr Med Res Opin 2006;22:1911–20.
[12]. Goswami R, Dufort P, Tartaglia MC, Green RE, Crawley A, Tator CH, Wennberg R, Mikulis DJ, Keightley M, Davis KD. Frontotemporal correlates of impulsivity and machine learning in retired professional athletes with a history of multiple concussions. Brain Struct Funct 2016;221:1911–25.
[13]. Hashmi JA, Baliki MN, Huang L, Baria AT, Torbey S, Hermann KM, Schnitzer TJ, Apkarian AV. Shape shifting pain: chronification of back pain shifts brain representation from nociceptive to emotional circuits. Brain 2013;136(pt 9):2751–68.
[14]. Haufe S, Meinecke F, Gorgen K, Dahne S, Haynes JD, Blankertz B, Biessmann F. On the interpretation of weight vectors of linear models in multivariate neuroimaging. Neuroimage 2014;87:96–110.
[15]. Hemington K, Rogachov A, Cheng J, Bosma R, Kim A, Osborne N, Inman R, Davis K. Chronic pain patients exhibit a complex relationship triad between pain, resilience, and within- and cross-network functional connectivity of the default mode network. PAIN 2018 In press.
[16]. Hemington KS, Wu Q, Kucyi A, Inman RD, . Abnormal cross-network functional connectivity in chronic pain and its association with clinical symptoms. Brain Struct Funct 2016;221:4203–19.
[17]. Hutchison RM, Womelsdorf T, Allen EA, Bandettini PA, Calhoun VD, Corbetta M, Della Penna S, Duyn JH, Glover GH, Gonzalez-Castillo J, Handwerker DA, Keilholz S, Kiviniemi V, Leopold DA, de Pasquale F, Sporns O, Walter M, Chang C. Dynamic functional connectivity: promise, issues, and interpretations. Neuroimage 2013;80:360–78.
[18]. Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM. FSL. Neuroimage 2012;62:782–90.
[19]. Krzywinski M, Altman N. Multiple linear regression. Nat Methods 2015;12:1103–4.
[20]. Kucyi A. Just a thought: how mind-wandering is represented in dynamic brain connectivity. Neuroimage 2017. doi: . [Epub ahead of print].
[21]. Kucyi A, Davis K. The neural code for pain: from single-cell electrophysiology to the dynamic pain connectome. Neuroscientist 2017;23:397–414.
[22]. Kucyi A, Davis KD. The dynamic pain connectome. Trends Neurosci 2015;38:86–95.
[23]. Kucyi A, Moayedi M, Weissman-Fogel I, Goldberg MB, Freeman BV, Tenenbaum HC, Davis KD. Enhanced medial prefrontal-default mode network functional connectivity in chronic pain and its association with pain rumination. J Neurosci 2014;34:3969–75.
[24]. Kucyi A, Salomons TV, Davis KD. Mind wandering away from pain dynamically engages antinociceptive and default mode brain networks. Proc Natl Acad Sci U S A 2013;110:18692–7.
[25]. Kutch JJ, Ichesco E, Hampson JP, Labus JS, Farmer MA, Martucci KT, Ness TJ, Deutsch G, Apkarian AV, Mackey SC, Klumpp DJ, Schaeffer AJ, Rodriguez LV, Kreder KJ, Buchwald D, Andriole GL, Lai HH, Mullins C, Kusek JW, Landis JR, Mayer EA, Clemens JQ, Clauw DJ, Harris RE. Brain signature and functional impact of centralized pain: a multidisciplinary approach to the study of chronic pelvic pain (MAPP) network study. PAIN 2017;158:1979–91.
[26]. Liegeois R, Laumann TO, Snyder AZ, Zhou J, Yeo BTT. Interpreting temporal fluctuations in resting-state functional connectivity MRI. Neuroimage 2017;163:437–55.
[27]. Loggia ML, Chonde DB, Akeju O, Arabasz G, Catana C, Edwards RR, Hill E, Hsu S, Izquierdo-Garcia D, Ji RR, Riley M, Wasan AD, Zurcher NR, Albrecht DS, Vangel MG, Rosen BR, Napadow V, Hooker JM. Evidence for brain glial activation in chronic pain patients. Brain 2015;138(pt 3):604–15.
[28]. Lutkenhoff ES, Rosenberg M, Chiang J, Zhang K, Pickard JD, Owen AM, Monti MM. Optimized brain extraction for pathological brains (optiBET). PLoS One 2014;9:e115551.
[29]. Miller EK, Cohen JD. An integrative theory of prefrontal cortex function. Annu Rev Neurosci 2001;24:167–202.
[30]. Napadow V, Kim J, Clauw DJ, Harris RE. Decreased intrinsic brain connectivity is associated with reduced clinical pain in fibromyalgia. Arthritis Rheum 2012;64:2398–403.
[31]. Power JD, Barnes KA, Snyder AZ, Schlaggar BL, Petersen SE. Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion. Neuroimage 2012;59:2142–54.
[32]. Ranganathan V, Gracey E, Brown MA, Inman RD, Haroon N. Pathogenesis of ankylosing spondylitis—recent advances and future directions. Nat Rev Rheumatol 2017;13:359–67.
[33]. Sacchet MD, Ho TC, Connolly CG, Tymofiyeva O, Lewinn KZ, Han LK, Blom EH, Tapert SF, Max JE, Frank GK, Paulus MP, Simmons AN, Gotlib IH, Yang TT. Large-scale hypoconnectivity between resting-state functional networks in unmedicated adolescent major depressive disorder. Neuropsychopharmacology 2016;41:2951–60.
[34]. Seeley WW, Menon V, Schatzberg AF, Keller J, Glover GH, Kenna H, Reiss AL, Greicius MD. Dissociable intrinsic connectivity networks for salience processing and executive control. J Neurosci 2007;27:2349–56.
[35]. Seminowicz DA, Moayedi M. The dorsolateral prefrontal cortex in acute and chronic pain. J Pain 2017;18:1027–35.
[36]. Treede RD, Rief W, Barke A, Aziz Q, Bennett MI, Benoliel R, Cohen M, Evers S, Finnerup NB, First MB, Giamberardino MA, Kaasa S, Kosek E, Lavand'homme P, Nicholas M, Perrot S, Scholz J, Schug S, Smith BH, Svensson P, Vlaeyen JW, Wang SJ. A classification of chronic pain for ICD-11. PAIN 2015;156:1003–7.
[37]. van der Linden S, Valkenburg HA, Cats A. Evaluation of diagnostic criteria for ankylosing spondylitis. A proposal for modification of the New York criteria. Arthritis Rheum 1984;27:361–8.
[38]. Varoquaux G, Raamana PR, Engemann DA, Hoyos-Idrobo A, Schwartz Y, Thirion B. Assessing and tuning brain decoders: cross-validation, caveats, and guidelines. Neuroimage 2017;145(pt B):166–79.
[39]. Wang D, Buckner RL, Fox MD, Holt DJ, Holmes AJ, Stoecklein S, Langs G, Pan R, Qian T, Li K, Baker JT, Stufflebeam SM, Wang K, Wang X, Hong B, Liu H. Parcellating cortical functional networks in individuals. Nat Neurosci 2015;18:1853–60.
[40]. Wu Q, Inman RD, Davis KD. Neuropathic pain in ankylosing spondylitis: a psychophysics and brain imaging study. Arthritis Rheum 2013;65:1494–503.
[41]. Xu Q, Yaksh TL. A brief comparison of the pathophysiology of inflammatory versus neuropathic pain. Curr Opin Anaesthesiol 2011;24:400–7.
[42]. Yarkoni T, Westfall J. Choosing prediction over explanation in psychology: lessons from machine learning. Perspect Psychol Sci 2017;12:1100–22.
[43]. Yeo BT, Krienen FM, Sepulcre J, Sabuncu MR, Lashkari D, Hollinshead M, Roffman JL, Smoller JW, Zollei L, Polimeni JR, Fischl B, Liu H, Buckner RL. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J Neurophysiol 2011;106:1125–65.
[44]. Zalesky A, Fornito A, Cocchi L, Gollo LL, Breakspear M. Time-resolved resting-state brain networks. Proc Natl Acad Sci U S A 2014;111:10341–6.
[45]. Zou H, Hastie T. Regularization and variable selection via the elastic net. J R Statist Soc B 2005;67:301–20.
Keywords:

Ankylosing spondylitis; Pain; Dynamic functional connectivity; Machine learning

© 2018 International Association for the Study of Pain