Chronic pain as a brain imbalance between pain input and pain suppression

Abstract Chronic pain is pain that persists beyond the expected period of healing. The subjective experience of chronic pain results from pathological brain network interactions, rather than from persisting physiological sensory input of nociceptors. We hypothesize that pain is an imbalance between pain evoking dorsal anterior cingulate cortex and somatosensory cortex and pain suppression (i.e. pregenual anterior cingulate cortex). This imbalance can be measured objectively by current density ratios between pain input and pain inhibition. A balance between areas involved in pain input and pain suppression requires communication, which can be objectively identified by connectivity measures, both functional and effective connectivity. In patients with chronic neuropathic pain, electroencephalography is performed with source localization demonstrating that pain is reflected by an abnormal ratio between the dorsal anterior cingulate cortex, somatosensory cortex and pregenual anterior cingulate cortex. Functional connectivity demonstrates decreased communication between these areas, and effective connectivity puts the culprit at the dorsal anterior cingulate cortex, suggesting that the problem is related to abnormal behavioral relevance attached to the pain. In conclusion, chronic pain can be considered as an imbalance between pain input and pain suppression.


Introduction
Over one-third of the world's population suffers from persistent or chronic pain, resulting in tremendous burden for the individual (Jensen et al., 2014) and society (Gaskin and Richard, 2012). The subjective experience of chronic pain results from pathological brain network interactions, rather than from persisting physiological sensory input via nociceptors (Jensen et al., 2014). Pain is processed by three known pathways (Rainville et al., 1997;Price, 2000;Fields, 2004;Eippert et al., 2009;Kong et al., 2010;. The two ascending pathways include the anatomically and functionally separable medial and lateral pain pathway (Price, 2000;Bushnell et al., 2013). The medial pain pathway, which involves the dorsal anterior cingulate cortex (dACC) and anterior insula as main hubs, encodes the unpleasantness and suffering of pain (Rainville et al., 1997;Price, 2000;Bushnell et al., 2013). The lateral pathway, which involves the somatosensory cortex (SSC) as major hub, processes the discriminatory/sensory components of the pain (Flor et al., 1995;Bushnell et al., 2013). In addition, a descending pain inhibitory pathway (Fields, 2004) involves the rostral and pregenual anterior cingulate cortex (pgACC), the periaqueductal gray, hypothalamus and rostral ventromedial brainstem (Fields 2004;Eippert et al., 2009;Kong et al., 2010). This descending pathway is responsible for stress-mediated pain inhibition (Yilmaz et al., 2010), and placebo analgesia (Eippert et al., 2009).
It has been proposed that pain is the result of an imbalance between brain areas that process pain input and pain suppression , and that if pain input equals pain suppression there is no pain perception, but if pain input is larger than pain suppression this results in subjective pain perception . This imbalance can be computed by analyzing the current density of the different frequency bands of the electroencephalogram (EEG) spectrum, source localized to the pgACC, dACC and SSC .
The frequency bands of interest within the EEG spectrum are h, a and c band activity, based on the thalamocortical dysrhythmia model of pain (Llinas et al., 1999;Sarnthein et al., 2005;Stern et al., 2006;Cauda et al., 2009;Walton et al., 2010;De Ridder et al., 2011;Schulman et al., 2011;Sametsky et al., 2015;Vanneste et al., 2018). The thalamocortical dysrhythmia hypothesis suggests that somatosensory deafferentation leads to a thalamocortical column-specific decrease in information processing, which permits slowing down of resting-state thalamocortical activity from normal a to the h frequency range (Llinas et al., 1999(Llinas et al., , 2005Vanneste et al., 2018). Decreased input also results in a reduction of GABA Amediated lateral inhibition, inducing c band activity surrounding the deafferented thalamocortical columns Graphical Abstract 2 | BRAIN COMMUNICATIONS 2021: Page 2 of 13 (Llinas et al., 2005). This c band activity surrounding h activity is known as the edge effect (Llinas et al., 1999(Llinas et al., , 2005. An abnormal increase of h oscillations in chronic pain patients has been identified Stern et al., 2006). Other studies, however, did not observe abnormal h oscillations in chronic pain (Schmidt et al., 2012;, but only a slowing of the peak a frequency in chronic pain (de Vries et al., 2013). Furthermore, abnormal h oscillations might basically represent the non-specific slowing of a activity observed in many chronic neurological and psychiatric disorders (Vanneste et al., 2018).
It has been proposed that the h-c coupling in thalamocortical dysrhythmia represents pain-related c activity (Babiloni et al., 2002) nested on h as a carrier wave (Lisman and Jensen, 2013), analogous to what has been proposed for cognitive processing (Canolty et al., 2006). This c activity nested on h is also known in other pathologies such as tinnitus, depression, Parkinson's disease, addiction, ADHD, etc. (Coullaut-Valera et al., 2014;De Ridder et al., 2015;Kim et al., 2015Kim et al., , 2017Mumtaz et al., 2017;Vanneste et al., 2018).
A balance between pain input and pain suppression requires communication between the involved areas. Communication in the brain can be analyzed by functional connectivity, effective connectivity and cross-frequency coupling. Functional connectivity reflects correlated activity between two or more areas in the brain, effective connectivity adds directionality to the functional connectivity and cross-frequency coupling looks at how high frequency activity (b and c) is nested on low frequency activity (d, h, a).
This study wants to confirm or disprove the proposed brain imbalance model for pain using EEG to analyze source localized activity (¼current density) and connectivity. Based on the thalamocortical dysrhythmia model for pain the frequencies of interest are h, a and c, the regions of interest are the main hubs of the lateral pathway (¼SSC), the medial pathway (¼dACC) and descending pain inhibitory pathway (¼pgACC). We hypothesize that pain is associated with an increased ratio >1 of the current density of (dACCþSSC)/2ÂpgACC, that the functional and effective connectivity is disrupted in the h and a band, and the cross-frequency coupling abnormal in hc.

Methods and materials Participants
This study consists of 100 subjects (50 women and 50 men; 34-66 years of age, M ¼ 48.47; SD ¼ 11.04): 50 healthy control subjects, and 50 subjects with chronic pain. The healthy control group reported no history of neurological or neuropsychiatric disorders, and no pain. A pain specialist screened the pain patients for neuropathic pain related to deafferentation (i.e. peripheral nerve, root, or central tract lesions). Whereas traditionally chronic pain is defined as pain lasting for more than 3 months (Treede et al., 2015), the patients in this study suffered pain for more than one year.
The pain score (pain percept: how severe is your pain?) on a numeric rating scale was 5.80 (SD ¼ 2.57) and on the pain vigilance and awareness questionnaire (PVAQ) 47.01 (SD ¼ 9.41). The PVAQ is a measurement that assesses the preoccupation with or attention to pain and pain changes, and is associated with pain-related fear and perceived pain severity (Roelofs et al., 2003). The study was in accordance with the ethical standards of the Helsinki declaration (1964) and was approved by the institutional ethics committee (UZA OGA85). All relevant data are available from the authors on request.

Electroencephalogram
Recordings EEG data (Neuroscan, http://compumedicsneuroscan.com/) were obtained in a quiet room while each participant was sitting upright on a comfortable chair. The EEG was recorded with 64 electrodes in the standard 10-10 International placement. The impedances were checked to remain below 5 kX. Data recording was eyes-closed (sampling rate ¼ 1 kHz, band passed DC-200 Hz) and lasted approximately 5 min. The ground electrode was located at AFZ and the reference was located at the vertex. Participants were instructed not to drink caffeinated beverages one hour and alcohol 24 hours prior before recording to avoid alcohol-or caffeine-induced changes in the EEG stream (Volkow et al., 2000;Logan et al., 2002;Siepmann and Kirch, 2002). By monitoring both slowing of the alpha rhythm and appearance of spindles in the EEG stream we checked the alertness of participants to prevent possible enhancement of the h power due to drowsiness during recording (Moazami-Goudarzi et al., 2010). Our data were resampled to 128 Hz, band-pass filtered in the range 2-44 Hz, plotted and carefully inspected for manual artifact-rejection. We removed all episodic artifacts including eye blinks, eye movements, teeth clenching, body movement, or ECG artifact from the stream of the EEG. We computed the average Fourier cross-spectral matrices for frequency bands d (2-3.5 Hz), h (4-7.5 Hz), a (8-12 Hz), b (13-30 Hz) and c (30.5-44 Hz).

Source localization analysis
Standardized low-resolution brain electromagnetic tomography (sLORETA, available at http://www.uzh.ch/keyinst/ loreta.htm) is a method for functional imaging yielding standardized current density with zero localization error based on certain electrophysiological and neuroanatomical constraints (Pascual-Marqui, 2002). This method is developed to estimate the intracerebral sources generating the scalp-recorded electrical activity in the five frequency bands defined (Song et al., 2015). The sLORETA algorithm solves the inverse problem by assuming related orientations and strengths of neighboring neuronal sources that are represented by adjacent voxels. The solution space used in this study and associated lead field matrix are those implemented in the LORETA-Key software. By applying the boundary element method on the MNI-152 (Montreal Neurological Institute, Canada) sLORETa implements revisited realistic electrode coordinates (Jurcak et al., 2007) and the lead field (Fuchs et al., 2002) on a MNI-152 volume with 6239 voxels together with a size of 5 Â 5 Â 5 mm. The co-registration makes use of the correct translation from the MNI-152 space into the Talairach and Tournoux space.

Region of interest analysis
The log-transformed electric current density was calculated for each regions of interest (ROIs) and averaged across all voxels belonging to the ROI for the fivet frequency bands included in this study. The ROIs in the present study are the left and right SSC, the dACC, and pgACC. We did not differentiate between left and right dACC, pgACC due to their proximity to the midline, as due to volume conduction, laterality is harder to differentiate for areas close to the midline.

Support vector machine
We used the SVM program in the data-mining software Weka to perform all classification tasks (Waikato Environment for Knowledge Analysis version 3.7, developed by the University of Waikato Machine Learning Group, available at http://www.cs.waikato.ac.nz/ml/weka/) (Smith and Frank, 2016) We build a predictive models by learning from examples provided in user supplied datasets and used the default settings as the running parameters. Our dataset included for each participant the five frequency bands for each ROI. The criterion for correct classification was defined by subjects being assigned to the correct group (e.g. for the full model: disorder versus healthy). The classification method that was used, was a linear logistic regression-based classifier as the classification method. A 10-fold cross-validation was performed on the full dataset (see Vanneste et al., 2018 for more detail information). The measurements of model accuracy calculated by the k-fold cross-validation technique include the true-positive ratio (TPR), false-positive ratio (FPR), root mean squared error (RMSE), mean average error (MAE), and j-statistic (see Vanneste et al., 2018 for more detail information about this specific measures). In order to determine significance of the model accuracy for pain/control, averaged model accuracy was calculated through randomization of the data. This was done by taking the same dataset used to generate the pain model and randomly reassigning patient data as either pain or control. This randomized dataset was then used to generate a prediction model and to model accuracy values. This was done 100 times, and the resulting randomized model accuracy statistics were averaged across all trials.

Lagged phase coherence
Phase coherence between time series corresponding to different spatial locations are usually interpreted as indicators of functional connectivity. Pascual-Marqui (2007a,b) introduced measures of coherence and phase synchronization that takes into account only non-instantaneous connectivity. Because this method the confounding factor of volume conduction is removed. The lagged phase coherence between two sources can then be interpreted as the amount communication between the regions contributing to the ROI defined. Since the two components oscillate coherently with a lag, the crosstalk can be interpreted as information sharing (more detail information can be found in Vanneste et al., 2014).

Granger causality
Granger causality reflects information transfer strength also called effective connectivity. By quantifying how much the signal in the seed region is able to predict the signal in the target region we calculated the effective connectivity from one region to another (Granger, 1969;Geweke, 1982). Granger causality is defined as the logratio between the error variance of a reduced model, which predicts one time series based only on its own past values, and that of the full model, which in addition includes the past values of another time series. In this study, we look at the granger causality between the pgACC, dACC, and the left and right SSC.

Cross-frequency coupling
h-b, h-c and a-c coupling are proposed to be effective means of communication between cortically distant areas (Canolty et al., 2006). To verify whether this is present, cross-frequency coupling was calculated for the pgACC, dACC, and the left and right SSC. Phase-amplitude was computed by computing the time-series for the x, y and z components of the current for each ROI. Next, these were filtered in the h, a, b and c frequency band-pass regions. For each ROI and each frequency band, a principal component analysis was calculated, and the first component was retained for the h and c bands. A Hilbert transform was then calculated on the gamma component and the signal envelope retained. Finally, the Pearson correlation between the h/a component and the envelope of the b/c envelope was computed for each participant.

Statistical analysis Support vector machine
To compare the different outcome measures (correctly classified, incorrectly classified, TPR, FPR, ROC, j-statistic, RMSE and MAE) of the SVM learning approach, we applied a univariate ANOVA with the model (test versus random) as the independent variable and the outcome measures as the dependent variable.

Region of interest
We performed a MANOVA including the log-transformed current density for the pgACC, dACC and the left and right SSC as dependent variables and group (chronic pain patients versus control subjects) as independent variables. These areas were selected based on the a priori hypothesis and were confirmed by SVM. For each area, we analyzed only the frequencies that showed up when applying in the SVM learning approach. Based on the outcome, a simple contrast analyses were conducted to look at specific effects for a specific frequency and a specific area. Pearson correlations were calculated between the region of interest and the pain score for the frequency bands at the areas that showed up during SVM. This analysis was corrected for pairwise comparisons using a Bonferroni correction. In addition, we looked at the balance between the areas that are part of the ascending pathway (i.e. dACC and the left/right SSC) and the descending pathway (i.e. pgACC). We include only the frequency bands that correlated with the pain percept as measured with the visual analogue scale. As the dACC and SSC combine two current densities and the pgACC only one, the current density of the pgACC was doubled, as such, the balance was calculated as follows: hjpgACC If the balance was greater than one, then it suggests that the key areas involved in the ascending pathways are more active in comparison to the key area involved in the descending pathway. If the balance was less than one, then it suggests that the key areas involved in the ascending pathways are less active in comparison to the key area involved in the descending pathway. If the balance is similar to one, then the key areas involved in the ascending pathways are as active as the key area involved in the descending pathway. A univariate ANOVA was applied to compare the balance between the control subjects and the chronic pain patients. In addition, a Pearson correlation was applied between the pain percept and the balance score.

Lagged phase coherence or functional connectivity
A comparison was made between chronic pain patients versus control subjects for all frequency bands. The threshold of significance for a given lagged phase coherence can be found as described by Pascual-Marqui (2002, 2007a. Time-series of current density were extracted for different regions of interest using sLORETA. Power in all 6239 voxels was normalized to a power of 1 and log transformed at each time point. Region-of-interest values thus reflect the log transformed fraction of total power across all voxels, separately for specific frequencies. The regions of interest selected were the pgACC, dACC, and left and right SSC. Lagged phase synchronization/ coherence or functional connectivity contrast maps were calculated for all frequency bands: d (2-3.5 Hz), h (4-7.5 Hz), a (8-12 Hz), b (13-30 Hz) and c (30.5-44 Hz). The significance threshold was based on a permutation test with 5000 permutations. This methodology corrects for multiple testing (i.e. for the collection of tests performed for all voxels, and for all frequency bands).

Granger causality
We applied Granger causality analysis to look at the direction of functional connectivity (i.e. effective connectivity). For the h frequency, we performed a MANOVA including the Granger causality for the effective connectivity (pgACC!left SSC, left SSC!pgACC, pgACC!right SSC, and right SSC!pgACC) as dependent variables, and group (chronic pain patients versus control subjects) as independent variables. Based on the outcome, simple contrast analyses were conducted for the individual connections. Based on the outcome, simple contrast analyses were conducted for the individual connections. We calculated Pearson correlations between effective connectivity and pain perception for all connections that showed a significant group-level effect.

Cross-frequency coupling
We performed a MANOVA including h-c and a-c phaseamplitude coupling for the both the left and right SSC and a-b coupling for the dACC as dependent variables and group (chronic pain patients versus control subjects) as independent variables. Based on the outcome, simple contrast analyses were conducted to look at specific effects. A Pearson correlation was computed between h-c and a-c coupling for the left and right SSC for both the chronic pain patients and control subjects. Furthermore, a Pearson correlation was calculated between the phaseamplitude coupling for h-c and a-c for the SSC, the abcoupling for the dACC, and pain perception as well as the balance between h-c and a-c for the SSC.

Data availability
Data will be made available upon reasonable request.

Support vector machine learning
Using a SVM learning approach, chronic pain (n ¼ 50) can be objectively detected in comparison to pain-free control subjects (n ¼ 50), solely based on brain activity of the dACC (including a and b band), the SSC (including h and c band), and the pgACC (including h and a band) with an accuracy of 89.3%. The other frequency bands, namely delta and beta, were not selected by the model. A random model has an accuracy not higher than chance (TPR: 50%) (F ¼ 2593.09, P < 0.001) (Fig. 1a, b). The FRP was significantly (F ¼ 22772.37, P < 0.001) lower for the pain model (M ¼ 0.16, SD ¼ 0.02) in comparison to the random model (M ¼ 0.50%, SD ¼ 0.02). The area under the curve shows a significant effect (F ¼ 7281.34, P < 0.001), indicating a higher score for SVM differentiates between respectively chronic pain versus controls for the pregenual anterior cingulate cortex (pgACC), for the h and a frequency band, dorsal anterior cingulate cortex (dACC) for the a and b frequency band and for the left and right somatosensory cortex (SSC) for the h and c frequency band. (C-F) Activity changes in the areas obtained with SVM show significant increase in chronic pain patients in comparison to control subjects for the specific frequency band obtained with SVM. (G) Correlations between the pain intensity percept (red) and the activity for the pgACC, dACC, left and right SSC and correlations between the pain distress (black) and the activity for the pgACC, dACC, left and right SSC reveals. (H) The difference in balance between the pgACC at a, and the dACC at b plus the left and right SCC at c frequency band is significantly increased in patients with chronic pain in comparison to control subjects suggesting that the key areas (dACC and SCC) involved in the ascending pathways are more active in comparison to the key area involved in the descending pathway. (I)This imbalance correlates with the perceived pain intensity. ( †P < 0.10, *P < 0.05, **P < 0.01, ***P < 0.001). For the radar plots, effective connectivity areas underlined indicate a significant effect (blue for pain distress, red for pain intensity).  (Fig. 1a, b). For the areas individually, the maximum accuracy was 68.8% (dACC: 67.4%; the SSC: 68.8%; pgACC: 63.3%) and the accuracy for the dACC and SSC together was 79.3%, confirming that pain can better be explained by looking at the pain input and suppression areas together (89.3% accuracy), i.e. as a balance.
A correlation analysis was performed between the current density and pain severity/distress as measured by the pain vigilance and awareness questionnaire. This analysis revealed a significant positive correlation between the current density in a in the pgACC (a: R 2 ¼ 0.30, P < 0.001) and a and b in the dACC (a: R 2 ¼ 0.28, P < 0.001j b: R 2 ¼ 0.42, P < 0.001) indicating the higher the score for pain related distress, the higher the current density, or vice versa. No effect was obtained for left SSC (h: R 2 ¼ 0.05, P ¼ 0.32j c: R 2 ¼ 0.01, P ¼ 0.37) and right SSC (h: R 2 ¼ 0.07, P ¼ 0.07j) c: R 2 ¼ 0.001, P ¼ 0.87) and the pgACC (h: R 2 ¼ 0.02, P ¼ 0.26) (Fig. lg). The effects continued after correction for multiple comparison using a Bonferroni correction (P < 0.008).
The balance between the areas of interest (pgACC, dACC and SSC) is significantly different from healthy pain-free subjects (F ¼ 17.14, P < 0.001) and furthermore correlates with intensity of the pain (R 2 ¼ 0.13, P ¼ 0.012) (Fig. lh-i). No correlation was obtained between the balance between these significant areas and pain distress (R 2 ¼ 0.07, P ¼ 0.08). After correction for multiple comparisons using a Bonferroni correction the effects remained (P < 0.025).

Functional connectivity
Functional connectivity, based on lagged phase synchronization between the ROI brain areas demonstrates that chronic pain patients have decreased functional connectivity in the a band (t ¼ 3.94, P < 0.05) and more functional connectivity in the h band (t ¼ 4.81, P < 0.05) (Fig. 2). No effect was obtained for d, b and c bands.
In patients with chronic pain both the information that is sent from pgACC to the left (R 2 ¼ 0.11, P ¼ 0.042) and right SCC (R 2 ¼ 0.11, P ¼ 0.041) negatively correlates with pain perception, as measured by the NRS. These effects, however, did not remain after correction for multiple comparison using a Bonferroni correction, No effect was obtained for the information that is sent from the left SSC (R 2 ¼ 0.01, P ¼ 0.52) and right SCC (R 2 ¼ 0.03, P ¼ 0.11) to pgACC. For pain distress, no significant correlations were obtained for pgACC to the left (R 2 ¼ 0.0001, P ¼ 0.99) and right SCC (R 2 ¼ 0.001, P ¼ 0.91) or from left SSC (R 2 ¼ 0.02, P ¼ 0.88) and right SCC (R 2 ¼ 0.02, P ¼ 0.39) to pgACC.
For the a band, an overall effect for the effective connectivity as measured with Granger causality was identified between the pgACC, dACC and the SSC (F ¼ 4.70, P < 0.001). A post-hoc univariate analysis indicated a significant increase (F ¼ 5.81, P ¼ 0.018) in information flow from the dACC to the pgACC for pain patients (M ¼ 0.021, SD ¼ 0.011) in comparison to control subjects (M ¼ 0.015, SD ¼ 0.014). From the pgACC to the dACC, we found a significant decrease (F ¼ 22.48, P < 0.001) in information flow for pain patients (M ¼ 0.013, SD ¼ 0.008) in comparison to control subjects (M ¼ 0.022, SD ¼ 0.010). Furthermore, the analysis yielded a significant decrease for information flow from the pgACC to the left and right SCC for pain patients (left: M ¼ 0.006, SD ¼ 0.0008; F ¼ 11.94, P ¼ 0.001; right: M ¼ 0.007, SD ¼ 0.007; F ¼ 6.29, P ¼ 0.014) in comparison to control subjects (left: M ¼ 0.012, SD ¼ 0.008; right: M ¼ 0.010, SD ¼ 0.009). Univariate analyses revealed no effect for effective connectivity between the left SSC (F ¼ 0.14, P ¼ 0.71); and right SCC (F ¼ 0.037, P ¼ 0.85) to the pgACC. Also, for the left SSC (F ¼ 0.10, P ¼ 0.76); and right SCC (F ¼ 0.12, P ¼ 0.74) to dACC no effect was obtained. No effect was obtained for the dACC to the left SSC (F ¼ 0.06, P ¼ 0.82); and right SCC (F ¼ 0.05, P ¼ 0.83).
A correlation analysis revealed only a significant effect between pain intensity perception using the NRS and information flow from the dACC to the pgACC (R 2 ¼ 0.24, P < 0.001), indicating that the more pain was perceived by patients, the more information flowed from the dACC to the pgACC. No correlation was obtained for the pain percept and information flow from the pgACC to the dACC (R 2 ¼ 0.01, P ¼ 0.90), from the pgACC to the left SSC (R 2 ¼ 0.02, P ¼ 0.58) or from the pgACC to the right SSC (R 2 ¼ 0.01, P ¼ 0.52), respectively. For pain distress, only a significant correlation was obtained between pain distress and information flow from the dACC to the pgACC (R 2 ¼ 0.19, P ¼ 0.002). No correlation was obtained for the pain distress and information flow from the pgACC to the dACC (R 2 ¼ 0.02, P ¼ 0.39), from the pgACC to the left SSC (R 2 ¼ 0.01, P ¼ 0.50) or from the pgACC to the right SSC (R 2 ¼ 0.01, P ¼ 0.53), respectively. The effect remained after correction for multiple comparison using a Bonferroni correction.

Phase amplitude coupling
For h-c coupling, a significant difference was obtained for the left and right SSC (F ¼ 7.68, P < 0.001) between (D) A correlation analysis between pain intensity perception the left and right SCC to pgACC demonstrates that more pain is related to less effective connectivity in chronic pain patients. (E-H) For the a frequency band, a decrease in effective connectivity from the pgACC to the left and right SSC as well as the dorsal anterior cingulate cortex (dACC) was found for chronic pain patients. In contrast, no difference in effective connectivity from the left and right SCC to the dACC to the pgACC was noted for chronic pain patients in comparison to controls subjects. (I) A positive correlation was identified between pain perception with effective connectivity from dACC to pgACC, indicating the higher pain intensities and distress correlate with increased effective connectivity from dACC to pgACC in chronic pain patients. No effect was obtained for the effective connectivity from pgACC to SCC (left and right) or from pgACC to dACC. ( †P < 0.10, *P < 0.05, **P < 0.01, ***P < 0.001). For the radar plots, effective connectivity areas underlined indicate a significant effect (blue for pain distress, red for pain intensity). Figure 4 Cross-frequency coupling in pain. A significant decrease in a-c coupling (a) and a significant increase in h-c coupling (b) was obtained for the left and right SCC for patients with chronic pain in comparison to control subjects. For the dACC, an increase in a-b coupling (c) was found for patients with chronic pain in comparison to control subjects. ( †P < 0.10, *P < 0.05, **P < 0.01, ***P < 0.001).
pain patients and control subjects. A post-hoc univariate analysis, revealed for right SSC (M ¼ 0.020, SD ¼ 0.012; F ¼ 14.64, P < 0.001) increased h-c coupling for pain patients in comparison to control subjects (M ¼ 0.012, SD ¼ 0.008). For the left SCC (M ¼ 0.017, SD ¼ 0.008; F ¼ 2.91, P ¼ 0.091), there was only a marginally significant increase in h-c coupling for the pain patients in comparison to control subjects (M ¼ 0.012, SD ¼ 0.009 (Fig. 4a-c)).
No correlation was obtained between the pain intensity percept and h-c coupling for the left SSC (R 2 ¼ 0.03, P ¼ 0.25) and right SSC (R 2 ¼ 0.002, P ¼ 0.79). For pain distress, no correlation was obtained with h-c coupling for the left SSC (R 2 ¼ 0.01, P ¼ 0.87) and right SSC (R 2 ¼ 0.001, P ¼ 0.90).
For a-b coupling, a significant effect SSC (F ¼ 10.22, P < 0.001) was obtained of the dACC when comparing pain patients (M ¼ 0.018, SD ¼ 0.009) in comparison to control subjects (M ¼ 0.013, SD ¼ 0.008). No correlation was obtained between the pain percept and h-c coupling for (R 2 ¼ 0.001, P ¼ 0.95) or between pain distress and h-c coupling (R 2 ¼ 0.001, P ¼ 0.80) for the dACC.

Discussion
This study provides converging evidence that chronic pain is an imbalance between brain areas involved in ascending and descending pain pathways. This imbalance is controlled by different brain oscillations that all contribute in routing of information flow between brain areas involved in pain input and pain suppression (Ploner et al., 2017).
This study confirms that chronic pain is associated with increased h and c oscillations in the SSC, the main hub of the lateral 'painfulness' pathway, in agreement with the thalamocortical dysrhythmia model (Llinas et al., 1999;Vanneste et al., 2018), and specifies c band activity in the somatosensory cortex correlates with intensity of chronic pain, similar to what is noted in acute experimental pain (Babiloni et al., 2002).
The dACC, as main hub of the medial 'suffering' pathway, accelerates from h to a and b frequencies, also suggesting an increase in pain processing.
The descending pain inhibitory system with the pgACC as main hub, which normally oscillates at rest in h (Vanneste et al., 2019). The results of this study demonstrate that the descending pain inhibitory systemaccelerate from h to a suggesting that pain suppression is paradoxically increased, possibly as an attempt, albeit insufficient, to compensate for the increase in pain input. This strongly suggests that indeed a balance better explains pain than only pain input.
The imbalance can improve either by reducing pain input or increasing pain suppression. Spinal cord stimulation seems to rely on increased pgACC activity (Moens et al., 2012), and by increasing pain suppression rebalancing pain suppression with pain input Vanneste, 2016, 2017). This not only seems to be the main mechanism of action of spinal cord stimulation, but of motor cortex stimulation (Lima and Fregni, 2008). But also opioids (Fields, 2004) and SNRIs such as duloxetine (Lunn et al., 2009;Kremer et al., 2018) exert their pain suppressing effects by modulation of the descending pain inhibitory system. Furthermore, if the descending pain inhibitory system is deficient, as in fibromyalgia, people perceive generalized pain .
A balance, by definition, requires communication between pain input and suppression between the dACC, SSC, and pgACC. Thus, the lack of communication, as demonstrated by decreased functional connectivity between pain input and pain suppression, results in aloss of balance between interacting areas. Indeed, the results of this study indicate functional connectivity changes between dACC, SSC, and pgACC for both the h and a band. A closer look at the data by analyzing effective connectivity indicates that for the h band, less information is sent from the pgACC to the left and right SSC. This finding is associated with increased pain and fits with the idea that the imbalance is due to reduced pain suppression. For a, increased information is sent from the dACC to the pgACC, which correlates with the intensity of the pain percept. This latter finding could suggest that the dACC is inhibiting activity in the pgACC, explaining the reduced information being sent from the pgACC to the dACC, and the left and right SSC. Conceptually, the brain attaches a paradoxical behavioural significance to the pain percept, by which it remains conscious. Indeed fluctuations of the salience network results in conscious perception of somatosensory stimuli at threshold (Boly et al., 2007). Furthermore, it has been shown that all psychosurgical interventions ultimately result in modulation of the pgACC (Schoene-Bake et al., 2010). Thus a dACC cingulotomy will reduce the dACC part of the ratio, possibly normalizing the imbalance by releasing the pgACC from inhibitory activity . And cingulotomies have been successfully used to treat chronic pain (Wilkinson et al., 1999;Cetas et al., 2008;Yen et al., 2009).
The findings of this study also identify increased h-c coupling in the SSC, which is consistent with the thalamocortical dysrhythmia model of pain that proposes that a-c nesting is a physiological mechanism transmitting sensory information and that slowed down alpha into the h range reflects pathological h-c coupling, associated with tinnitus and pain (Llinas et al., 1999;De Ridder et al., 2015).
The balance disorder also suggests that there are multiple ways the brain can generate pain, either related to too much pain input, or because of deficient pain suppression, or a combination of the two. This could benefit a new brain-based classification of different types of pain, rather than the traditional etiology-based categorization of pain and may help to develop a mechanism-based approach for selecting therapy. Indeed, fibromyalgia pain has been attributed to a deficiency in pain suppression  and a balance problem between pain input and pain suppression (De Ridder and Vanneste, 2017), and this imbalance can be normalized by occipital nerve field stimulation (De Ridder and Vanneste, 2017). It is of interest that this imbalance mechanism might be more universal in view of the pathophysiological analogy between pain, tinnitus, Parkinson's disease and major depression (Llinas et al., 1999;Vanneste et al., 2018) as well as reward-deficiency disorders, such as obesity, addiction, hyperactivity and personality disorders (Blum et al., 2014). Indeed, the described mechanism involves the same brain areas involved in reward-deficiency disorders, resulting from allostasis, i.e. reference resetting .
Furthermore, the fact that pain can be detected by machine learning with 89.3% accuracy, and if this concept is generalizable, will permit to develop objective measures for other subjective states such as tinnitus, depression, anxiety by measuring ratios of current density in the hubs that drive the ascending and descending pathways, rather than measuring electrical activity of only input areas of sensory pathways as was routinely done in for example tinnitus (van der Loo et al., 2009), pain (Babiloni et al., 2002) and depression (Jaworska et al., 2012). The imbalance theory also permits to more rationally prescribe medications by clinicians. For example, combining medication that activate pain inhibition, with medications that inhibit pain input will be more effective in rebalancing pain or tinnitus or depression states than medication that only addresses one of the pain pathways. Thus, the magic bullet concept of Nobel laureate Paul Ehrlich (Winau et al., 2004) does not hold in this concept, except if one medication would simultaneously modulate all three pathways. Ehrlich envisioned that just like a bullet fired from a gun aiming to hit a specific target, there must be a pharmacological way to specifically target a symptom or disease, in his case invading microbes. A more rational pain treatment however is to combine a drug that activates the descending pain inhibitory pathway, with another one that suppresses the medial pathway and a third that suppresses the lateral pathway. This cocktail approach is similar to what is very successfully used in for example AIDS pharmacotherapy, keeping the disease under control in 85-90% of patients (Lu et al., 2018).
Although chronic pain is one of the most important medical problems facing society, there has been limited progress in finding an objective measure for this fundamentally subjective state. A conceptually new way of defining pain electrophysiologically may aid in developing better subtyping, better diagnostic measures and better treatments for chronic pain.
In conclusion, this paper presents a fundamentally new concept of pain as an imbalance disorder in the brain, which may have large implications, not only as a basis for finding an objective measure for a subjective pain state, but also for developing better pain medication, novel neurostimulation designs, and subtyping pain. Furthermore, in view of the analogy of the underlying pathophysiology of pain, tinnitus, depression, Parkinson's disease (Vanneste et al., 2018), and slow wave epilepsy (Llinas et al., 1999), there is no reason to believe this concept could not be extended to other subjective states.

Funding
No funding.