Nine female participants (47.59 ± 16.54 years, 8 right:1 left-handed) with musculoskeletal pain who did not self-report misusing opioids were enrolled in a 4-week group ACT intervention program (Table 1). Participants were referred from the outpatient practice of a physician certified in Physical Medicine and Rehabilitation with subspecialty certification in Pain Medicine. The practice involves rehabilitation and management of both acute pain and CMP, with a higher prevalence of females versus male patients reporting CMP (Meier et al., 2021).

Participants were required to be at least 18 years of age, speak English, have been living with musculoskeletal CP for 3 or more months, have a Brief Pain Inventory (BPI) Score of ≥4, and have no history of cancer or malignancy, head or severe body trauma in the past 6 months. Participants with neurologic (e.g., history of stroke, brain lesions, or intracranial surgery) or psychiatric disorders not commonly comorbid with CP were excluded. Patients who were not addicted to opioids but were taking opiates on a PRN (“as needed”) basis were eligible to participate, in order to reflect real-world clinic conditions as closely as possible. Only one participant self-reported using PRN opioid medication. Most participants reported having more than one type of chronic pain. Specifically, four patients reported suffering from fibromyalgia, one reported Ehlers–Danlos syndrome, one reported Conradi–Hünermann syndrome, two reported chronic neck pain, three reported chronic lower back pain, and one reported trigeminal neuralgia (for further details regarding the participants’ medical histories, please see Supplementary Table 4).

Patients completed two 90-min manualized ACT sessions per week for 4 weeks (Luoma et al., 2007; Potter, 2012). ACT sessions were administered by two licensed, trained Certified Therapeutic Recreation Specialists (CTRS/L). Behavioral outcomes were measured using validated assessments including the Acceptance and Action Questionnaire (AAQ-II) (Bond et al., 2011), the Center for Epidemiologic Studies Scale (CES-D) (Radloff, 1977; Vilagut et al., 2016), the Chronic Pain Acceptance Questionnaire (CPAQ) (McCracken et al., 2005; McCracken and Vowles, 2006; Vowles et al., 2008; Vowles and Thompson, 2011), the NIH Toolbox Neuro-QoL^TM (Quality of Life in Neurological Disorders) scales (Cella et al., 2012), and the NIH Patient-Reported Outcome Measurement Information System (PROMIS) measures of pain interference (Amtmann et al., 2010, 2011), administered using an iPad (see Supplementary Material). The study protocol (#6991) was approved by the University of New Hampshire Institutional Review Board on July 19, 2018. All participants provided informed consent.

The behavioral assessment data were entered into Excel spreadsheets using Qualtrics software (Qualtrics, 2005) for data management. Statistical analyses were conducted using SAS^® v.9.4. (SAS Institute Inc., 2019). Paired Student’s t-tests and Wilcoxon signed rank tests were used to assess differences in behavioral measures from pre-to-post ACT (subtracting post minus pre scores). Positive or negative change scores indicated satisfactory results, depending on the specific test in question (e.g., reduced CES-D scores indicated improvements in depression while higher AAQ scores indicated improvements in pain acceptance).

MRI data were collected before and immediately after 4 weeks of ACT using a Siemens Three Tesla Magnetom Prisma scanner at Boston University, Boston, MA, United States. Structural MPRAGE was collected (TR/TE = 2.53 s/1.32 ms, flip angle = 7°, field of view (FOV) = 256 × 320 mm, 0.8 mm³ resolution) to allow for anatomical registration. Subsequently, two runs of 8 min resting-state fMRI data were obtained using a T2^∗ weighted Echo Planar Imaging sequence (2.5 mm³ resolution, 60 slices, TR/TE = 1.2 s/30 ms, 300 volumes, FOV = 205 mm, multi-slice interleaved ascending) for all participants. During the resting state scans, participants were instructed to lie still in the scanner with their eyes open, fixating on a crosshair placed in their field of view. Only the first of the two resting state scans were used for analysis. Two rsfMRI scans were collected in the case that one set was unusable (e.g., movement artifact). The first scan set was of high enough quality to use.

Standard preprocessing steps were carried out using Statistical Parametric Mapping, version 12 (SPM12, Penny et al., 2006). First, all scan data were imported in the form of DICOM images and converted to Nifti files using the DICOM Import function in SPM12. Functional data were realigned and co-registered to the standard Montreal Neurological Institute (MNI) template in SPM12. Motion correction, band-pass filtering (0.0078–0.08 Hz), slice-timing correction, normalization to remove individual variability for between subject comparisons, and smoothing to increase signal to noise ratio were carried out using SPM12 (Figure 1, step 1). Next, each participant’s brain was parcellated into discrete regions of interest representing nodes from the Power atlas (Power et al., 2011) using Mango (Multi-image Analysis GUI; Lancaster et al., 2010). The mean time course within seed regions were extracted from the residual images using REX (Duff, 2008) (Figure 1, step 2). FC estimates were then calculated using the pairwise Pearson correlation of seed regions located time course across all 264 nodes resulting with a 264 by 264 connectivity matrix (Figure 1, step 3). Finally, connectivity matrices were reduced to a subset of 101 nodes that were associated with the FPN, DMN, and SN.

Graph analysis applied to fMRI is a powerful way of characterizing brain networks. In this context, a network represents a collection of nodes, and the functional connections between pairs of nodes. Nodes in large-scale brain networks represent brain regions, with connections being anatomical, functional, or effective, depending on the type of imaging data analyzed. Application of graph theory-based approaches have identified biologically plausible brain networks found to topologically organize in a non-trivial manner (e.g., network integration and modular structure) and support efficient information processing of the brain (Bullmore and Sporns, 2009; Rubinov and Sporns, 2010; Wang et al., 2010; Bullmore and Bassett, 2011). These network analyses allow us to visualize the connectivity pattern across the entire brain and also quantitatively characterize its global organization (Wang et al., 2010; Sporns, 2018). In the current study, we leverage the ability of graph analytics to identify network connections linking rsfMRI connectivity to ACT treatment.

We used the Network Based Statistic (NBS; Zalesky et al., 2010) to examine FC in the DMN, SN, and FPN. This graph theory-based method provides a statistical approach to identify changes in FC associated with diagnostic status or changing psychological contexts (Zalesky et al., 2010). The NBS is based on the principles underpinning traditional cluster-based thresholding of statistical parametric maps. We use the NBS to identify significant network connectivity differences between pre-ACT and post-ACT (Figure 1, step 4). We tested for within network connectivity changes for each of the three networks of interest independently, and then we tested for network connectivity changes across all 101 nodes that make up the FPN, DMN, and SN. Results presented represent functional network differences for t > 2.5 (10,000 permutations). We further examine whether specific pairwise connections in brain networks affected by ACT are related to behavior, and whether connectivity changes associated with ACT correlate with changes in behavior/outcome measures (using Pearson’s r).

Causal mediation analyses were conducted, with confounder adjustment, to evaluate associations between FC changes and behavioral changes (pre- to post-ACT). In other words, we tested whether connectivity changes reflected by the NBS (Figure 1, step 4) were mediators of changes in depression, social role, and CP acceptance scores. A counterfactual mediation approach was implemented in this analysis using SAS PROC CAUSALMED (Valeri and VanderWeele, 2013; VanderWeele, 2015, 2016). Although structural equation modeling (SEM) is frequently used to examine mediation (Baron and Kenny, 1986), its limitations include the following: (1) It does not have a clear theoretical foundation for defining causal mediation effects; (2) It does not deal with confounding and interaction effects effectively; (3) It does not treat binary outcomes and binary mediators in a unified manner. The counterfactual approach overcomes these limitations (MacKinnon et al., 2007; SAS Institute Inc., 2019).

Causal mediation analysis quantifies and estimates the total, direct, and indirect (or mediated) effects between an independent variable and an outcome. It enables causal interpretations of these effects under the assumptions of the counterfactual framework (Robins and Greenland, 1992; Pearl, 2011). The causal mediation model decomposes the total effect into a direct effect [e.g., the effect of an independent variable [A] on outcome [Y; A = 0 vs. A^∗ = 1] and the natural indirect effect (NIE)]. The controlled direct effect (CDE) simulates a randomized controlled trial (RCT) by hypothetically assigning the same value of the mediator to all individuals in the sample, with bootstrapped standard errors (Naimi et al., 2014). The NIE captures the effect of the mediation pathway (e.g., the average change in Y if the exposure is fixed to the level of the intervention and the mediator changes accordingly [e.g., A = 0 to A^∗ = 1]; VanderWeele, 2015). The mediation path is represented by arrows “B” and “C” in Figure 2.

For this analysis, we used the theoretical constructs underlying ACT to guide our approach (e.g., that psychological flexibility, in the context of CP, includes factors pertaining to acceptance and cognitive defusion (learning to experience uncomfortable thoughts, feelings, and sensations in a way that reduces their interference on valued activities and roles in one’s daily life); Chin and Hayes, 2017). Thus, we focused on indicators of CP acceptance, pain interference, depression, and social roles as key behavioral outcomes.

Significant changes in FC from pre- to post-ACT in the DMN, SN, and FPN were observed. Using the NBS, within network effects of ACT were only observed in the SN which consisted of a sub-network of four nodes and three functional connections (Figure 3A, t > 2.5, p = 0.039). No effects of ACT were observed within the DMN or FPN. NBS tests comparing pre-ACT vs. post-ACT of the triple network (DMN, FPN, and SN nodes combined) identified a network of 10 nodes and 10 connections displaying decreases in FC associated with completing ACT (Figure 3B, t > 2.5, p = 0.05). Interestingly, the within network ACT effects observed in the SN were also present in the triple network. Between network changes were also observed in the triple network where all DMN and FPN nodes connected to SN nodes (Figure 3B).

Next, we assessed the relationships between brain connectivity changes and behavioral outcomes using Pearson correlation statistics. Six of the ten functional connections that showed significant differences in strength within the Triple Network from pre- to post-ACT (shown in Figure 3B) were significantly correlated with behavior change scores. Pearson correlations and corresponding p-values for the following significant relationships are shown in Table 3. Results indicated that ACT effects in connectivity (functional connections from NBS analyses shown in Figure 3) between the MTG and aCC (52, −59, 36; −11, 26, 25) and between the PFC and aCC (38, 43, 15; 0, 30, 27) were correlated with social role (Neuro-QoL^TM) and acceptance (AAQ-II) scores. FC changes between the rIPL and aCC (two functional connections connecting 44, −53, 47 with −11, 26, 25 and 0, 30, 27) were correlated with Neuro-QoL^TM depression scores; and changes between rSMG and aCC (55, −45, 37; 10, 22, 27) were correlated with both depression and social role scores. Changes between rSMG and aCC (55, −45, 37; −11, 26, 25) and rIPL and aCC (44, −53, 47; −11, 26, 25) were correlated with reduced pain severity scores.

In the causal mediation framework for this study, a baseline behavioral variable (e.g., Pre-ACT Depression Score) was hypothesized to relate to an outcome variable (Post-ACT Depression score) via the causal mechanism that is represented in Figure 2. As depicted in the diagram, the total causal treatment effect pertaining to the outcome (Post-treatment behavioral score) consists of the following two parts: (1) a direct effect; (2) a mediated (or indirect) effect via the “functional connectivity” variable representing change in FC between two brain regions.

Results from our exploratory causal mediation models (Table 4) demonstrated that improvements in specific behavioral outcome scores were significantly related to both the direct pathway “A” (e.g., baseline behavior scores predicted post-treatment scores) and through the indirect mediation pathway “B and C” via changes in FC (the NIE). Statistically significant NIE provide evidence in support of our mediation hypothesis for rSMG-aCC for depression, MTG-aCC for social role, both rSMG-aCC and MTG-aCC for CP acceptance, and rlPL-aCC for pain interference. These relationships persisted after covariate adjustment for age, BMI, pain severity, and other potential confounders. By contrast, statistically significant mediation effects were not observed for rlPL-aCC for depression, PFC-aCC and social role, or MTG-aCC and AAQ-II. No significant interactions were detected.

Several studies have evaluated the association between fMRI and behavioral outcomes in mindfulness-based interventions (e.g., MBSR), though the majority of these evaluations are not specific to ACT (Tang and Leve, 2016; Feliu-Soler et al., 2018; Zeidan et al., 2018). A recent study (Yoshino et al., 2018) used a CBT intervention with behavioral activation (an ACT component) and found that the OFC played an important role in improvements in pain intensity post-treatment. Activation of the dorsal pCC at pre-treatment was also associated with improvements of clinical symptoms. ACT has also demonstrated sustained medium-large effect sizes on social functioning (Dahl et al., 2004; Feliu-Soler et al., 2018).

Yu et al. (2019) examined ACT-oriented treatment for fatigue in 354 adults with CP. Pearson’s correlations and hierarchical regression were conducted to investigate associations between improvement in fatigue interference and improvements in outcomes. Mixed effects models were used to explore associations between baseline fatigue interference and changes in outcome measures. Results suggested that participants improved in perceptions of fatigue interference, pain, psychological flexibility (PF) processes, and daily functioning. Changes in fatigue interference were associated with changes in pain, PF processes, and daily functioning | r| = 0.20–0.46. Changes in fatigue interference were associated with changes in pain acceptance independent of changes in pain perception. The authors concluded that individuals with fatigue appeared to benefit from the ACT−oriented interdisciplinary treatment for CP, and relatively higher levels of fatigue did not appear to decrease the treatment benefit. Although fatigue was not a focus in our study, the results are similar in terms of demonstrating that the behavioral improvements associated with ACT may persist regardless of pain severity and the presence of other co-morbidities.

Smallwood et al. (2013) examined gray matter volume (GMV) differences between CP patients and healthy controls and found that there were 12 clusters where GMV was decreased in CP patients compared with controls. These clusters included many regions that are considered part of the “pain matrix” involved in pain perception, but also included many other regions that are not commonly regarded as pain-processing areas. The authors also reported that the most common behavioral domains associated with these regions were cognitive, affective, and perceptual domains, suggesting that many of the regions may relate to the constellation of comorbidities that often accompany CP (e.g., fatigue, depression, cognitive, and emotional impairments).

Using causal mediation analysis to assess whether changes in connectivity strength mediated changes in specific behavioral outcomes, we observed statistically significant mediation effects for rSMG-aCC with depression, MTG-aCC with social role, rSMG-aCC and MTG-aCC with CP acceptance, and rlPL-aCC with pain interference. Because the models were adjusted for age, BMI, pain severity, and other behavioral covariates, we were able to determine that the relationships were not confounded by these factors.

We also observed significant mediation effects for rSMG-aCC and rlPL-aCC with pain severity, despite the fact that changes in perceived pain severity are not considered direct targets of ACT. In our unadjusted analyses, perceived pain severity scores did not change significantly from pre- to post-ACT, yet the controlled direct effect in the causal mediation models, with confounder adjustment, demonstrated significant changes in both direct and indirect effects. These exploratory analyses underscore the complexity of measuring pain perception, particularly as other therapeutically targeted behavioral changes and associated neural connectivity changes may be occurring simultaneously.

Notably, the median baseline CES-D score among our participants was 16, indicative of high depressive symptomatology (Vilagut et al., 2016) concurrent with CP. After the ACT intervention, the median CES-D score was reduced to 7, and this change appears to be mediated by decreased rSMG-aCC hyperconnectivity.

By contrast, statistically significant mediation effects were not observed for rlPL-aCC with depression, MTG-aCC with AAQ-II, or PFC-aCC with social role. In these cases, the relationships were confounded by other factors and may operate via more complex multiple mediation pathways that could not be tested in this small exploratory sample. For example, although we observed a statistically significant mediation effect for PFC-aCC and social role in unadjusted models, inclusion of body mass index (BMI), pain severity, and baseline depression nullified this relationship.

A growing body of literature has begun to employ mediation analysis to assess the mechanisms underlying behavioral and clinical outcomes (Atlas et al., 2014; DasMahapatra et al., 2015; Gu et al., 2015; Lee et al., 2015; Cederberg et al., 2016; Sanders et al., 2017; Mischkowski et al., 2019; Whibley et al., 2019). However, few studies have assessed behavioral changes with respect to fMRI data using causal mediation analysis, particularly with respect to CP. Lindquist (2012) described an extension of SEMs applied to data from a fMRI study of thermally induced pain. The results suggested that many classic “pain-responsive regions” such as the anterior insula showed significant mediation of the temperature-induced relationship, and that subjective pain ratings increased near the end of the stimulation period. Other regions, such as the insular cortex appeared to be active during pain judgment. Like our study, this study supports the use of mediation modeling in future research to better understand how connectivity changes among different brain regions affect the subjective experience of pain, and may inform pain management approaches.

In addition to the small sample size, the limitations of our study include the quasi-experimental research design and the lack of long-term follow-up data beyond the immediate post-ACT period. A randomized controlled trial would have provided more methodological rigor to the study, and randomized trials with longer follow-up periods are needed to confirm the findings reported here.

Villanueva et al. (2019) investigated three key aspects of ACT, including its effectiveness, long-term follow-up, social context, and social processes. The authors contend that researchers should include longer follow up periods in clinical studies (Gloster et al., 2013). This is especially important for treatment-resistant patients (e.g., patients who do not respond to standard, first line treatments), for whom viable treatment options are limited (Gloster et al., 2020). The variable number of years that participants suffered from chronic pain is also a limitation in our study, although all participants meet the clinical definition of chronic pain lasting more than 6 months. The heterogeneity of the participants’ pain conditions also precludes inference about how ACT affects specific types of chronic pain. Due to the comorbid pain conditions experienced by the participants in our sample, we are unable to evaluate the results of ACT on any single type of musculoskeletal pain.

However, the mechanism of ACT may transcend specific genetic and acquired diagnoses that result in chronic pain, which clinicians may perceive as a strength of the intervention. Central sensitization (the enhanced activity of neurons and circuits in nociceptive pathways) underlies many of the changes in chronic pain (Latremoliere and Woolf, 2009), illustrating the contribution of the central nervous system to the generation of pain hypersensitivity across many types of pain diagnoses. Because of central sensitization, chronic pain is not necessarily coupled to the original stimuli source. Moreover, the abnormal pain sensation and sensitivity resulting from central sensitization is enhanced by stress. Aligning with the findings from our study, prior researchers (Lewis et al., 2012) have highlighted neuroplastic changes in the insula (within the salience network and the DMN) with respect to processes affecting mood disturbances, working memory, fatigue, and body awareness issues observed across different chronic pain conditions. More recently, Tang et al. (2020) found that disruptions of multiple networks, including the DMN, salience network, and limbic system, may contribute to the neurophysiological mechanisms underlying postherpetic neuralgia. A growing body of research also suggests that central sensitization may be driven by neuroinflammation in the peripheral and central nervous system, causing widespread chronic pain. For example, Ji et al. (2018) reported that sustained increase of cytokines and chemokines in the central nervous system may promote chronic pain that affects multiple body sites. These processes warrant further study in larger randomized controlled trials.

The fact that participants were all female also limits the generalizability of our findings. Although both males and females were recruited and referred for study participation, the study protocol required daytime commitment to therapy and out-of-state travel. Males cited inflexible work schedules as the main reason why they were unable to participate in the study. However, it is also important to note that prior researchers have documented that the prevalence of chronic pain is higher among females than males (Tsang et al., 2008). Researchers have reported that in addition to psychosocial characteristics that may vary between males and females, differences in genotype and endogenous opioid functioning play a causal role in these disparities, and considerable literature implicates sex hormones as factors influencing pain sensitivity (Bartley and Fillingim, 2013). However, the specific modulatory effect of sex hormones on pain among men and women requires further exploration. Thus, clinicians must consider a myriad of factors when developing a treatment approach to chronic pain, including gender.

Lastly, omitted variable bias may also pose a limitation in our study. For example, factors outside therapy itself, including social processes, may account for up to 33% of improvement in patients undergoing psychotherapy and group-based interventions (Cuijpers et al., 2012). It remains poorly understood how the influence of social surroundings longitudinally affects patients’ well-being, social function, and pain perception (Biglan and Embry, 2012; Wilson et al., 2014). Prior research suggests that both close and extended social ties may be relevant for positive outcomes (Kuehner and Huffziger, 2013). Additional research is needed to better understand the variation in treatment outcomes in relation to an individual’s social and environmental context, using an exposome lens (Juarez et al., 2014). Future research should also consider different ways of measuring pain perception and should evaluate both mediators and moderators of ACT in pain as well as in other chronic diseases.