Neural synchrony in the pain connectome predicts chronic pain severity and interactions with mental health outcomes: A transdiagnostic study using magnetoencephalography and multivariate modeling

Participants

We recruited n = 106 active-duty military personnel and Veterans of the Canadian Armed Forces (CAF), as part of a cross-sectional study on the relationships between brain, cognitive, and mental health outcomes in military. After quality assurance procedures, we retained n = 99 usable sets of individual data (mean age = 46.8 years, SD = 9.8 years; 88% male, n = 87). From the 106 initial sets, three were removed due to the inability of the participants to complete either MEG or MRI scans for various reasons, such as scanner malfunction, or participants not being able to fit in the scanner(s). In addition, two sets were removed during preprocessing, due to excessive sensor noise. Another three participants were removed due to incomplete questionnaire data.

The participants served in a wide variety of capacities across numerous roles including infantry, armoured, ordnance disposal, special operations forces, paratroopers, artillery, and administrative or technical roles. Inclusion criteria were being between the ages of 20 and 65 and the ability to be imaged in an MRI scanner’s confined space. Exclusions were made for individuals with ferrous metal implants that could affect MEG or MRI scanning, or those with a history of seizures, or active substance use disorders. Written informed consent was obtained from all participants, and the Hospital for Sick Children Research Ethics Board approved the study protocol. All participants provided their written informed consent for study participation.

Data Collection

Five minutes of eyes-open resting-state MEG data (600 Hz) in the supine position was collected using a 151-channel CTF system at the Hospital for Sick Children in Toronto. Fiducial coils, placed at the left and right pre-auricular points as well as the nasion, were used for continuous head motion monitoring. For MEG source modelling, MRI scans were collected on a 3T Siemens PrismaFit scanner equipped with a 20-channel head and neck coil. A high-resolution 3D MPRAGE T1-weighted image was obtained with an isotropic voxel size of 0.8 mm (TR of 1870 ms, TE of 3.1 ms, and TI of 945 ms) over a field of view measuring 240 × 256 mm and comprising 240 slices each 0.8 mm thick.

On the day of scanning, participants completed self-reported screening questionnaires for psychiatric, neurological, neurobehavioural, neurocognitive, and pain symptoms. These included the McGill Pain Questionnaire [44], the Generalized Anxiety Disorder (GAD-7) scale [45], the Patient Health Questionnaire 9 (PHQ-9) screener [46], and the PTSD Checklist-Military Version (PCL-M) [47], which is based on DSM-IV criteria to maintain consistency with prior research.

MEG Signal Processing

MEG data were processed using FieldTrip and SPM toolboxes [48,49]. Data were bandpass filtered (1-150 Hz), and epoched into 10-second pseudo-trials and the maximum number of trials were retained after removing trials where head position deviated more than 5 mm from the original position. In addition, epochs including SQUID jumps in excess of ±2 pT were also removed. Subsequently, cardiac and ocular artefacts were removed following independent component analysis (ICA) using the “fastica” method from FieldTrip, and visual inspection of the resulting components.

Source localization was performed to pinpoint cortical activity within thirty-six regions of interest (ROIs) within the pain connectome [12,13]. To localize sources, a forward solution was created for each participant by using the individual anatomical T1-weighted MRI scans. After co-registration of the MEG data to the MRI, a linearly constrained minimum variance (LCMV) beamformer [50] was employed to estimate electrophysiological activity at the predetermined ROIs, using the epoched, artifact-free sensor-level data.

For connectivity analysis, the broadband, source-level time-series data were bandpass filtered into the canonical frequency bands: delta (1-3 Hz), theta (4-7 Hz), alpha (8-14 Hz), beta (15-30 Hz), low-gamma1 (30-55 Hz), low-gamma2 (65-80 Hz), and high-gamma (80-150 Hz). The Weighted Phase-Lag Index (wPLI) was used to quantify functional connectivity between pairwise ROI connections [51]. A 36-by-36 connectivity matrix was constructed for each frequency band and epoch per participant by calculating the wPLI between all possible ROI pairs. These matrices were subsequently averaged across epochs to yield a functional connectivity matrix per participant for each frequency band.

We provide the ROIs’ coordinates in the Supplementary Material.

Partial Least Squares Modeling

We employed PLSR [37,38] to model the complex relationships between neural synchrony as independent variables and specific outcomes, including pain severity, the interaction of pain severity with anxiety symptoms, the interaction of pain severity with depressive symptoms, and the interaction of pain severity with PTSD symptoms, as dependent variables. This technique is well-suited to highly collinear data and cases where the number of observations is smaller than the number of predictors. Prior to conducting the PLSR analysis, we controlled the effects of sex and age using multiple linear regression. Our analysis included separate models for each dependent variable: pain intensity, pain-anxiety interaction, pain-depression interaction, and pain-PTSD interaction across each frequency band.

For statistical inference of the model coefficients, we utilized a resampling procedure. Specifically, we performed a permutation test by resampling the subject labels 1,000 times and rerunning the PLSR on each resampled data. This process generated a distribution of coefficients from the resampled models corresponding to each original coefficient. We then compared the coefficients from the actual model to this distribution to calculate a p-value for each coefficient, thereby assessing their statistical significance.

This approach underscored the role of neural synchrony in predicting pain intensity and its interactions with mental health conditions.

Predictive Modeling

Significant connections identified through PLSR modeling served as input variables. To predict pain and its interactions with various mental health conditions, we employed leave-one-out cross-validation (LOOCV). Five regressor algorithms were used: elastic net, random forest, partial least squares, support vector machine with a linear kernel, and support vector machine with a radial kernel. Model performance was compared using mean absolute error (MAE) and R² for goodness of fit. The algorithm with the lowest MAE was selected to generate actual-predicted plots. Our target variables included pain, and pain-anxiety, pain-depression, and pain-PTSD interactions.

Modeling and Visualization Tools

We used the R programming environment and packages “pls” and “caret” for statistical modeling [52–54], along with “ggplot2” and “BrainNet Viewer” for visualization [55,56].

Chronic pain severity moderately correlates with mental health outcomes

The severity of chronic pain, anxiety, depression, and PTSD were positively skewed (Figure 1A), highlighting the need for non-parametric statistical methods. Spearman correlations between chronic pain and mental health outcomes were generally low to moderate (all p-values were significant after multiple comparisons with medium effect sizes: p_{pain & anxiety} = 3.92e-4, p_{pain & depression} = 1.11e-4, p_{pain & depression} = 1.31e-4) (Figure 1B). Conversely, pairwise correlations between mental health outcomes were high (all p- values significant after multiple comparisons with large effect sizes).

• Download figure
• Open in new tab

Figure 1. Severity measures and their associations.

(A) Histogram of pain, anxiety, depression, and PTSD severities across subjects. (B) Spearman’s correlation matrix of chronic pain and mental condition measures. All correlations are statistically significant after multiple comparison correction. Asterisks denote correlation coefficients with large effect sizes.

The medium effect sizes observed between pain and mental health outcomes indicated that, while these associations were related, they were not entirely overlapping, supporting further exploration into how interactions between pain and specific mental health outcomes including anxiety, depression, and PTSD might relate to neural measurements. Despite the high intercorrelation among mental health outcomes, examining the neural correlates of pain’s interaction with each mental health condition could yield distinct insights, capturing the complexity of these comorbid conditions within the pain connectome.

Chronic pain is associated with neural synchrony in the pain connectome and interacts with mental health outcomes

PLSR modelled the association between chronic pain and network neural synchrony, and the interaction between pain and mental health outcomes. The PLSR method is suitable for collinear and nonparametric contexts, effectively projecting the complexity of the data in a robust and reliable manner [37,38] and models the association between continuous clinical measures and their interaction with functional connectivity.

Chronic pain severity is associated with frequency-specific neural synchrony

Significant associations between chronic pain and neural synchrony were identified across many connections in the pain connectome, particularly in the beta and high-gamma frequency bands (Figure 2, Row 1). Theta synchrony exhibited an inverse relationship with pain across three connections in the somatosensory and subcortical networks, contrary to the direct association of the left amygdala to the medial prefrontal cortex. Beta synchrony showed large-scale significant effects with pain, exhibiting both positive and negative associations, regions involving the default mode and somatosensory networks being anti-correlated with pain intensity, and connections with the central executive and subcortical networks showing the greatest number of positive correlations with pain. Across the low-gamma ranges, several within-network and between-network connections were anti-correlated with pain, contrary to the connection from the left primary somatosensory cortex to the left medial temporal lobe, which is directly related to pain severity in low-gamma2 band. In the high-gamma band, connections from regions involving the central executive network, the default mode network, and subcortex to other regions were mostly correlated with pain intensity, while other regions were both correlated and anti-correlated with pain severity.

• Download figure
• Open in new tab

Figure 2. Transpectral neural synchrony is associated with pain severity and significantly interacts with mental health outcomes.

Brain maps show connections with their regional synchronization intensity significantly associated with pain and the interaction between pain and mental health across various frequency bands, along with the strength of this relationship, obtained through PLSR. Nodes represent regions from the pain connectome and are color-coded by network membership. The colors of the links denote the direction of the association; red indicates a direct relationship between connection strength and clinical measure, while blue indicates an inverse relationship. The width of the connections also indicates the strength of the effect. Abbreviations: CE: Central Executive, DM: Default Mode, LIM: Limbic, SAL: Salience, SM: Somatomotor, SC: Subcortical, VA: Ventral Attention, VIS: Visual.

Neural synchrony reliably predicts chronic pain severity and the interaction of pain and mental health outcomes

We used the significant connections identified in the previous section across all frequency bands to perform leave-one-out cross-validation for predicting pain severity and the interactions of pain with mental health. Figure 3 illustrates the performance of predictive models for the dependent variables using various machine learning methods.

• Download figure
• Open in new tab

Figure 3. Transpectral neural synchrony reliably predicts chronic pain severity and pain-mental health interactions.

Each column demonstrates the prediction of a target variable. The top barplots show the Mean Absolute Error (MAE) in the predicting of target variables by a set of algorithms. The middle barplots display the R-squared (R²) values as the measure of goodness of fit for the predictions by each algorithm. The bottom row scatter plots indicate the actual versus predicted values for the best-performing algorithm with the lowest MAE. The reported correlation values are Spearman’s, and the axes represent the ordered actual and predicted values. (Abbreviations: EN: ElasticNet, RF: Random Forest, PLS: Partial Least Square, LSVM: Linear Support Vector Machine, RSVM: Radial Support Vector Machine)

For the prediction of pain severity, partial least squares machine learning method yielded the minimum absolute error (MAE = 9.45), which corresponds to 16.57% of the maximum score on the pain severity scale. It also achieved the highest goodness of fit (R² = 0.40), with a Spearman’s correlation of 0.70 between the actual and predicted values. In predicting the pain-anxiety interaction, the partial least squares algorithm also proved to be the best predictive algorithm, achieving MAE = 148, along with R² = 0.38 and a Spearman’s correlation of 0.48 between the actual and predicted values. Regarding the interaction of pain and depression, elastic net was the most predictive algorithm, exhibiting a slightly lower error compared to partial least squares which also had an MAE of 179, R² = 0.27, and a Spearman’s correlation of 0.46. For the prediction of pain-PTSD interaction, elastic net also demonstrated the best performance, with MAE = 516, R² = 0.36, and a Spearman’s correlation of 0.55 between the actual and predicted values.

Summary

In this study, we comprehensively examined the complex relationship between neural synchrony and chronic pain in the pain connectome, as well as the neural synchrony associated with the interaction between chronic pain and mental health conditions such as anxiety, depression, and PTSD symptoms (Figure 2).

Using PLSR applied to MEG-based synchrony, we uncovered significant associations between chronic pain severity and neural synchrony across multiple frequency bands, with the most pronounced effects occurring in the beta and high-gamma bands, across various brain networks, with minor but significant associations in theta and low-gamma frequency bands.

Our results reveal that chronic pain is not only linked to specific patterns of neural synchrony but also that cross-frequency connections are influenced by co-occurring mental health symptoms. The interaction between pain and anxiety was notably mediated by neural synchronization in theta low- (theta) and high-frequency (gamma) ranges, while the interaction between pain and depression, and pain and PTSD, were associated with synchronisation primarily in the gamma band, but also a small number of connections in the beta band. These findings highlight the complex repertoire of multispectral neural states that underlie the percept of chronic pain and show co-occurring mental health challenges significantly interact with the neural architecture supporting pain percepts.

Overall, these results show that transpectral synchrony within specific brain networks is a predictive biomarker for both pain severity and interactions with mental health (Figure 3). This provides a potential pathway for developing personalized biomarkers and interventions in chronic pain management.

Comparisons with prior work

Comparing our results on chronic pain mechanisms with existing literature on MEG spectral power analysis within the pain connectome reveals both consistency and contradictory results in the neural substrates of chronic pain (Figure 2, Row 1). For example, Kisler et al. reported decreased low-gamma power in the default mode and ascending nociceptive pathway [16], which partially aligns with our findings of minor salience and default mode connections negatively associated with pain in this band. However, their observation of increased alpha power contrasts with our lack of significant connectivity effects in the alpha band – but of course, regional, oscillatory power changes are distinct mechanisms from network-level phase synchronisation [57]. Kim et al. (2019) found increased alpha and decreased beta band power in chronic multiple sclerosis pain [18], which are consistent with those connections negatively associated with pain in the beta band and contrast with other beta connections and the lack of significant alpha connections in our observations. Fauchon et al.’s exploration of peak alpha frequency abnormalities in chronic pain conditions [20], with sex-based differences, offers a compelling comparison to our results. We regressed out sex from the analysis and did not find significant connections for the alpha synchronisation, although we did not look at peak alpha frequency. Diers et al. demonstrated reduced beta activity during conditioned pain modulation, linking beta oscillations to top-down pain regulation [21], and Jin et al. showed increased beta activity in the somatosensory cortex during attentional focus on pain, highlighting beta’s role in cognitive control [22]. These findings align with our result, where beta synchrony shows both positive and negative associations with pain intensity. Another study by Kim et al. (2020) examining cross-network functional coupling [23] revealed abnormalities in alpha, beta, and gamma frequency bands in chronic pain, resonating with our findings of significant connectivity associations in various frequency bands. Studies by Parker et al. and Lim et al. on spectral power shifts in response to pain or interventions [58,59] align with our connectivity patterns in similar frequency bands but contrast with our emphasis on connectivity over spectral power. Choe et al., Hsiao et al., and Iwatsuki et al. provide direct comparisons, highlighting decreased global connectivity and frequency-specific reorganization in chronic pain, reinforcing the importance of examining neural connectivity [60–62]. Dinh et al.’s exploration of increased connectivity at theta and gamma frequencies in frontal brain areas of chronic pain patients [63] intersects with our findings of high connectivity in the theta and high-gamma bands. Their emphasis on frontal connectivity as a target for interventions aligns with our observations.

Our observed pain-anxiety interactions in the theta and gamma frequency bands aligns with broader literature implicating theta oscillations in cognitive and affective processes related to anxiety [64] and oscillations in sensory processing and pain perception [16,60] (Figure 2, Row 2). These findings suggest that connectivity in these bands is crucial in integrating and modulating pain and anxiety experiences. The anti-correlation between pain-anxiety interaction and subcortical beta synchronization points to the role of beta in modulating neural mechanisms underlying these conditions [21–23,65]. The absence of significant associations in the alpha and delta bands, despite alpha’s consistent implication in anxiety and pain [16–19,65], may hint at distinct oscillatory power and neural synchrony roles. This suggests that while alpha power alterations are a generalized marker of stress, connectivity patterns may not directly correspond to pain-anxiety interactions.

MEG studies in major depressive disorder (MDD) and chronic pain have highlighted nuanced neural connectivity patterns bridging emotional and physical pain processing. Nugent et al. reported no significant differences in functional connectivity between healthy controls and individuals with MDD [66]. However, other studies have identified specific connectivity patterns predictive of therapeutic outcomes, such as alpha connectivity between the dorsolateral prefrontal cortex and subgenual anterior cingulate cortex [67], and altered connectivity dynamics in MDD, particularly in beta and low-gamma bands [68–70]. Research in chronic pain has documented altered beta oscillations in modulation of pain and abnormalities in functional coupling across beta and gamma bands [21–23, 62], mirroring the disrupted connectivity patterns observed in depression. This consistency highlights a shared neural architecture between chronic pain and depression. Our findings align with both depression-specific and chronic pain observations, suggesting common neural processing alterations in both conditions, particularly in gamma and beta bands (Figure 2, Row 3). This convergence emphasizes the roles of these bands in the co-occurrence of pain and depression. The absence of significant connections in the alpha band in our pain-depression interaction results also contradicts weaker partial directed coherence observed in depression [71] and parallels peak alpha frequency abnormalities in chronic pain [20].

Comparing our results with existing literature on resting-state MEG connectivity and network measures for pain and PTSD enhances our understanding (Figure 2, Row 4). Studies show that beta and gamma frequency activity correlates with PTSD severity and altered neural connectivity [72]. Similarly, pain research highlights frequency abnormalities in chronic pain patients, particularly in the beta and gamma bands [21–23]. These studies suggest shared neurophysiological mechanisms linking pain and PTSD that are consistent with our results on the association of beta and gamma synchrony and pain-PTSD interaction. High-gamma band dynamics in PTSD reveal significant insights on high-frequency oscillations predicting PTSD severity and altered connectivity patterns [73,74]. These insights mirror our observations on pain-PTSD interactions in the high-gamma band. Theta band alterations in both PTSD and chronic pain are well-documented, with chronic pain studies observing increased theta connectivity and PTSD research variably reporting increased connectivity related to symptom severity [63,75]. Our findings do not represent theta connectivity’s role in pain-PTSD interaction.

Predictive modelling

Applying machine learning techniques to predict chronic pain, pain severity, and perception of pain has attracted attention in pain research [76,77]. Our study demonstrates the effectiveness of these algorithms in predicting pain severity based on neural synchrony with appropriate predictive accuracy. By using the entire frequency spectrum for modeling, we can capture a broader range of neural states, which enhances the performance of predictive modeling (Figure 3). Our result aligns with existing literature focusing on neuroimaging with varying degrees of success [78–84]. Some studies have reported high accuracy in pain prediction, often utilizing advanced algorithms; others have shown moderate predictive performance, underscoring the complexity of pain as a subjective and multifaceted phenomenon. The variability in predictive performance across studies can be attributed to several factors, including differences in data types (e.g., EEG vs. fMRI), the underlying neural mechanism being measured, the specific algorithm employed, and the pain phenotypes being modeled [85–87]. Moreover, despite the promising results in pain prediction, there remains a need for robust external validation of these models to ensure their applicability in diverse clinical settings. Some studies have successfully validated their models on independent datasets, thereby demonstrating their potential for clinical translation [84]. However, challenges such as inadequate reporting standards and potential biases persist, limiting the generalizability of these models across different populations and pain conditions, an issue that has been addressed in some research [88].

In the interaction between pain and mental health outcomes, our study extends the use of machine modeling approach using neural synchrony inputs to explore these complex relationships, particularly focusing on predicting the interaction of pain with anxiety, depression, and PTSD (Figure 3). The comorbidity of chronic pain with mental health disorders is well-documented, with evidence suggesting that the interplay between pain and mental health exacerbates both conditions, leading to increased disability and impairment [25]. Chronic pain often co-occurs with anxiety, depression, and PTSD, forming a bidirectional relationship where each condition intensifies the other [26,27,32,33]. Despite the extensive evidence on the comorbidity of pain and mental health disorders [25–36], employing machine learning to model these interactions using neuroimage data remains limited. Our study underscores the importance of this approach, suggesting that machine learning techniques can effectively capture the intricate relationships between chronic pain and mental health outcomes.

Limitations & Future Direction

There are several limitations to this study. The cross-sectional nature of the data limits our ability to infer causality between neural synchrony changes and chronic pain severity or its interaction with mental health. Longitudinal studies are needed to better understand these temporal dynamics and disentangle “state versus trait” pain percepts. Indeed, the establishment of registries such as the DoD’s Uniformed Services University Pain Registry Biobank represent efforts that allow the collection of multimodal pain-related data longitudinally, promising to add to our understanding of the developmental trajectory of chronic pain [89]. While our machine learning models show promise, the generalizability of these models is unknown without external validation using independent datasets and different pain populations would be critical to ensure applicability in clinical settings.

For future studies, several directions are recommended. Longitudinal studies should be conducted to explore the causal relationship between neural synchrony and the development or exacerbation of chronic pain and comorbid mental health conditions. Furthermore, integrating multimodal data, including genetic, behavioral, and other neuroimaging modalities, could enhance the accuracy of predictive models and lead to more personalized treatment strategies. Emphasizing personalized medicine, future interventions could be tailored to target specific neural synchrony patterns identified in individuals, potentially improving outcomes in chronic pain management. Finally, exploring the effects of therapeutic interventions, such as neurostimulation or mindfulness-based therapies, on neural synchrony within these specific frequency bands could provide new avenues for treating chronic pain, especially in patients with comorbid mental health disorders.

Conclusion

The findings of this study offer a deeper understanding of the neural complexity underlying chronic pain and its significant interactions with mental health conditions such as anxiety, depression, and PTSD. By employing advanced techniques such as partial least squares regression and machine learning, we identified that neural synchrony across specific frequency bands can serve as predictive biomarkers for both pain severity and its comorbidity with mental health disorders. These insights not only align with but also expand upon existing literature, highlighting the intricate and multifaceted nature of chronic pain and its connections to mental health, reinforcing the importance of not viewing chronic pain and mental health as independent conditions in clinical settings, nor in terms of the underlying neurobiological mechanisms. We emphasise the need for a multidisciplinary approach to the assessment and treatment of chronic pain, integrating chronic pain’s physical and mental health dimensions. This integrated approach is crucial for developing more effective, comprehensive care models that reflect the true complexity of chronic pain, its significant mental health comorbidities, and its neurobiology.