Resting-state electroencephalography and magnetoencephalography as biomarkers of chronic pain: a systematic review

Abstract

Reliable and objective biomarkers promise to improve the assessment and treatment of chronic pain. Resting-state electroencephalography (EEG) is broadly available, easy to use, and cost efficient and, therefore, appealing as a potential biomarker of chronic pain. However, results of EEG studies are heterogeneous. Therefore, we conducted a systematic review (PROSPERO CRD42021272622) of quantitative resting-state EEG and magnetoencephalography (MEG) studies in adult patients with different types of chronic pain. We excluded populations with severe psychiatric or neurologic comorbidity. Risk of bias was assessed using a modified Newcastle–Ottawa Scale. Semiquantitative data synthesis was conducted using modified albatross plots. We included 76 studies after searching MEDLINE, Web of Science Core Collection, Cochrane Central Register of Controlled Trials, and EMBASE. For cross-sectional studies that can serve to develop diagnostic biomarkers, we found higher theta and beta power in patients with chronic pain than in healthy participants. For longitudinal studies, which can yield monitoring and/or predictive biomarkers, we found no clear associations of pain relief with M/EEG measures. Similarly, descriptive studies that can yield diagnostic or monitoring biomarkers showed no clear correlations of pain intensity with M/EEG measures. Risk of bias was high in many studies and domains. Together, this systematic review synthesizes evidence on how resting-state M/EEG might serve as a diagnostic biomarker of chronic pain. Beyond, this review might help to guide future M/EEG studies on the development of pain biomarkers.

1. Introduction

Chronic pain is one of the most prevalent and cost-intensive health conditions worldwide.¹¹⁶ In this context, the inherently subjective nature of experiencing pain has long been recognized as a key issue when dealing with chronic pain in clinical and research settings.¹¹⁹ To improve assessment and guide treatment of chronic pain, the quest for objective measures of acute and chronic pain has gained increasing attention.¹⁴ For these purposes, the development of reliable and objective biomarkers of chronic pain has become a key challenge in pain research.⁶⁴ In line with the complexity of pain, there is an increasing awareness that there may not be a single, uniform biomarker but that different biomarkers subserving different functions, eg, for the diagnosis, monitoring, prognosis, and prediction of treatment responses may be needed to best capture pain.⁹

Over the past decades, alterations in brain structure and function, ranging from the molecular to the network level, have been increasingly recognized in patients with chronic pain.⁵³ Consequently, neuroimaging methods have been used to develop noninvasive, brain-based biomarkers of chronic pain. Electroencephalography (EEG) has gained attention as an easy-to-use and cost-efficient tool to assess brain function with a high temporal resolution in chronic pain. Electroencephalography is complemented by magnetoencephalography (MEG). Magnetoencephalography measures brain signals closely related to EEG. However, MEG is technically more demanding and less cost-efficient. Both methods have been used to investigate changes in oscillatory brain activity and functional connectivity at different frequencies as potential biomarkers of chronic pain. In particular, slowing of the peak frequency in the alpha band (8-13 Hz)^21,96 and increases of oscillatory brain activity at theta frequencies (4-8 Hz)^69,76,100 have been reported and discussed. However, M/EEG approaches are not standardized, and results are heterogeneous and inconsistent.

Here, we aimed at synthesizing the available evidence on resting-state M/EEG as biomarkers of chronic pain in adult humans. To this end, we conducted a systematic review of cross-sectional and longitudinal studies of patients with chronic pain using resting-state M/EEG as a measure of brain function. These studies can particularly serve the development of diagnostic, monitoring, and predictive biomarkers of chronic pain.

2. Methods

This review was conducted and is reported in accordance with the most recent Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines (PRISMA⁷⁷). The project protocol was registered on PROSPERO on September 10, 2021 (CRD42021272622). Deduplication of records, title and abstract screening, full-text review, and data extraction were conducted using Covidence software.¹²

2.1. Search strategy

The databases MEDLINE (through PubMed), Web of Science Core Collection (through Web of Science), Cochrane Central Register of Controlled Trials (through Ovid), and EMBASE (through Ovid) were searched on September 15, 2021. All searches were repeated before the final analysis on December 29, 2021. No historical limit was applied, and no filters for study type were used. For EMBASE, we excluded conference proceedings. In addition, we screened reference lists of included studies and personal files for further relevant publications.

The search string combined “electroencephalography” or “magnetoencephalography” and related terms with “chronic pain” and specifiers for several relevant chronic pain conditions, which do not entail “chronic pain” in their name (eg, fibromyalgia or postherpetic neuralgia). To this end, we searched the recently published International Association for the Study of Pain classification system of chronic pain conditions¹⁰⁸ for relevant diagnostic entities and terms. The complete search strategy is available in the supplementary material (available at https://links.lww.com/PAIN/B751).

2.2. Study selection

Study inclusion and exclusion criteria are presented in Table 1. In summary, we included peer-reviewed studies that used quantitative resting-state M/EEG to analyze brain activity in chronic pain conditions. We excluded studies in patients with primary headache conditions or severe psychiatric/neurologic comorbidity because abnormalities in resting-state M/EEG that are unrelated to chronic pain have been found in migraine,⁷⁸ severe depression,¹⁷ and neurodegenerative disease.⁹⁹

Table 1 - Inclusion and exclusion criteria.

Inclusion (included, if all apply)	Exclusion (excluded, if any applies)
Published, peer-reviewed study	Review article or case report
Chronic pain primary condition studied	Primary headache condition
M/EEG during resting state	Severe psychiatric or neurologic comorbidity
Reporting of quantitative M/EEG data
Human participants ≥18 y

M/EEG, magneto-/electroencephalography.

2.3. Record screening, full-text review, and data extraction

Two authors (P.T.Z., M.P.) screened titles and abstracts for eligibility blinded to each other's decisions. In case of disagreement, a third researcher (V.D.H.) was consulted, and conflicts were discussed. The same procedure was followed for full-text review. Data extraction was performed by one author (P.T.Z.) and checked by another author (V.D.H.). Extracted data comprised quantitative M/EEG measures (peak alpha frequency [PAF], frequency-specific power, frequency-specific connectivity), patient characteristics (sample size, sex, age, diagnostic entity, type of pain [nociceptive, neuropathic, nociplastic, mixed], pain intensity), and study design according to the study by Grimes and Schulz.³⁴

2.4. Data synthesis strategy

For data synthesis, we grouped the studies according to study designs. By doing so, we were able to relate study designs to different types of biomarkers according to the Biomarkers, EndpointS, and other Tools-initiative⁹ (see Table 2 for an overview of study and biomarker types). Cross-sectional studies (comparison of patients and healthy participants) can serve the development of diagnostic biomarkers of chronic pain. Longitudinal studies (longitudinal descriptive, randomized and nonrandomized studies) can help to develop biomarkers for the monitoring of chronic pain and, in few cases, for the prediction of treatment responses. Descriptive studies correlating M/EEG with pain intensity in a single assessment can be useful to establish biomarkers for the diagnosis and monitoring of chronic pain.

Table 2 - Study designs and biomarker types.

Study design	BEST biomarker(s)
Cross-sectional observational	Diagnostic
Longitudinal descriptive	Monitoring
Longitudinal randomized and nonrandomized controlled trial	Monitoring, predictive (treatment response)
Descriptive	Diagnostic/monitoring

BEST, Biomarkers, EndpointS, and other Tools⁹

Due to the reported outcome measures and the high level of heterogeneity of study designs, a formal meta-analysis was not feasible. For instance, for cross-sectional comparisons of theta power, 47% of all included studies, 79% of studies with negative results, and 23% of studies with positive results did not report the required parameters for meta-analysis. Hence, formal meta-analysis would have introduced a reporting bias. Therefore, we used vote-counting and modified albatross plots³⁶ for semiquantitative data synthesis. By plotting P values against sample sizes for different directions of effects, albatross plots allow for graphically estimating effect sizes for studies with similar research questions (eg, is there a difference in alpha power between patients with chronic pain and healthy participants?). However, because included studies used heterogeneous statistical methods for hypothesis testing, we did not superimpose effect size estimation contours on the plots. Modified albatross plots were used for comparison of PAF, frequency-specific power, and frequency-specific connectivity at theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz), and gamma (30-80 Hz) frequencies in cross-sectional (diagnostic biomarker), longitudinal (monitoring biomarker), and descriptive studies (monitoring/diagnostic biomarker). For graphical presentation of longitudinal study results, we only included studies that reported pain relief. In case of multiple P-values for different regions of interest in a single study, we reported the lowest P value. P values were displayed as reported in the primary studies, independently of possible adjustment for multiple comparisons. In case of imprecise reporting of P values (eg, “P < 0.05”), we chose the nearest decimal (eg, “P = 0.049”) for graphical representation in albatross plots.

When summarizing results across studies for a certain parameter (eg, theta power in cross-sectional studies), we focused on vote counting and labeled results as positive, if more studies were found for either the “lower” or “higher” category compared with the “nonsignificant” and the respective other category.

For studies reporting quantitative M/EEG measures other than the aforementioned variables (eg, microstate analysis or machine learning algorithms) and/or using M/EEG as a predictive biomarker, narrative data synthesis was performed due to the low overall number of studies and high heterogeneity of methods and outcome measures.

2.5. Risk of bias and quality assessment

Risk of bias and study quality were assessed using a modified version of the Newcastle–Ottawa Scale¹¹⁷ adapted by Pinheiro et al.⁸⁷ for EEG studies on pain. This tool assesses risk of bias and quality of studies included in systematic reviews and/or meta-analyses across the domains “selection of study participants” (4 items), “comparability/confounders” (2 items), and “outcome data” (3 items). Although in the original version stars are awarded for single domains, we rated items as “high” (negative for study quality) or “low” (positive for study quality) risk of bias because this allows for easier interpretation of scoring results. We did not calculate sum scores for single studies because single items had frequently to be scored “n/a” (not applicable), and comparison of sum scores across studies would have been misleading.

Furthermore, we evaluated all included studies regarding adherence to core open science principles (preregistration, sample size calculation, correction for multiple testing, and sharing of primary data and code).

Risk of bias and quality assessment was performed by one author (P.T.Z.) and checked by another author (V.D.H.).

3. Results

3.1. Study selection

Searching databases resulted in 2430 results after deduplication. We added 2 more studies from citation screening and personal records. Following title and abstract screening, 141 studies were identified for full-text review. Finally, 76 studies were included.^{1,5–8,10,11,15,16,18–27,30–33,35,37–49,51,56,59,60,63,66–69,74–76,79,80,82,89–91,95–98,100–105,109,110,112–115,120,122,124–126}Figure 1 depicts the PRISMA flow diagram of study selection along with reasons for study exclusion at every stage.

Figure 1.: PRISMA flow diagram of study selection. CP, chronic pain; EEG, electroencephalography; PRISMA, preferred reporting items for systematic reviews and meta-analyses; RS, resting state; qM/EEG, quantitative M/EEG.

3.2. Study characteristics

Figure 2 and Table 3 summarize the main characteristics and results of included studies. Thirty-two studies were conducted in patients with mixed pain, 26 in patients with neuropathic pain and 18 studies in patients with nociplastic pain (mainly fibromyalgia). Most studies used EEG (n = 66), whereas only few studies used MEG (n = 10). The most frequent study type was cross sectional (n = 41; comparing patients to healthy participants), followed by longitudinal descriptive studies (n = 17; longitudinal tracking of patients). Five studies comprised different study designs, ie, cross-sectional comparison of patients and healthy participants at baseline and longitudinal tracking of patients. Longitudinal studies included controlled and noncontrolled interventional designs and were heterogeneous regarding the type of intervention (Table 3). The most frequent interventions were noninvasive neuromodulation, cognitive–behavioral therapy, surgery, and pharmacological approaches. Sample sizes varied between 10 and 342 participants (median 38.0, see Fig. S1, https://links.lww.com/PAIN/B751 for a graphical illustration). In cross-sectional studies, these sample sizes included not only patients but also the sum of patients and healthy participants. Studies were published between 1999 and 2021 (Fig. S2, https://links.lww.com/PAIN/B751).

Figure 2.: Pain type, recording modality, and study designs of included studies. The charts show absolute numbers of studies. EEG, electroencephalography; MEG, magnetoencephalography.

Table 3 - Characteristics of included studies.

Author and year	Study design	Intervention	EEG/MEG	Condition	Type of pain	Total sample size	Pain duration (mo)	Pain intensity (0-10)	Main results (EEG/MEG analysis)
Ahn et al. 2019 1	(Randomized) controlled trial	tACS vs sham	EEG	Back pain	Mixed	20	84.8	4.4	Negative correlation between somatosensory alpha power and pain intensity Higher somatosensory alpha power after alpha-tACS compared with sham tACS Negative correlation between stimulation-induced alpha power increase and pain relief
Barbosa-Torres and Cubo-Delgado, 2021 5	Longitudinal descriptive	NFB	EEG	Fibromyalgia	Nociplastic	37	n.rep.	8.4	Higher SMR/theta wave ratio after NFB Lower pain after NFB
Baroni et al. 2020 6	Cross-sectional observational	—	EEG	Orofacial pain	Mixed	48	49.2	6.4	No power differences between patients and controls
Bell et al. 2001 7	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	31	57.6	n.rep.	Higher alpha power (posterior midline) in patients compared with controls
Bernardi et al. 2021 8	Longitudinal descriptive	tACS vs tRNS	EEG	Fibromyalgia	Nociplastic	36	n.rep.	n.rep.	Higher alpha power after tACS compared with tRNS Lower pain symptoms after tACS compared with tRNS
Camfferman et al. 2017 10	Descriptive	—	EEG	Mixed	Mixed	103	n.rep.	n.rep.	Correlation between alpha and theta power and pain intensity
Choe et al. 2018 11	Cross-sectional observational	—	MEG	Fibromyalgia	Nociplastic	37	36.6	6.1	Lower theta global connectivity and connectivity (within the default mode network/between middle/inferior temporal gyrus and visual cortex) in patients compared with controls Correlation between pain duration and reduced connectivity (between inferior temporal gyrus and visual cortex)
Day et al. 2021 16	Longitudinal descriptive	Different cognitive and/or behavioral interventions	EEG	Low back pain	Mixed	57	168	4.6	Lower theta, alpha, beta power across all interventions, no differences between the groups Positive correlation between beta power reduction (in central ROI) and pain intensity reduction in mindfulness-based cognitive therapy group only Lower pain across all interventions
Day et al. 2020 15	(Randomized) controlled trial	Different cognitive and/or behavioral interventions	EEG	Low back pain	Mixed	69	173.5	4.8	Moderation analysis: baseline alpha and theta power did not moderate changes in pain
De Melo et al. 2020 19	(Randomized) controlled trial	tDCS (different protocols) vs sham	EEG	Fibromyalgia	Nociplastic	31	79.2	6.7	Lower frontal and parietal alpha power after 5d tDCS Lower pain across all interventions
De Melo et al. 2021 18	Descriptive	—	EEG	Fibromyalgia	Nociplastic	31	79.8	6.7	No correlation between alpha power and pain intensity
De Ridder and Vanneste, 2017 20	Cross-sectional observational	tDCS vs sham	EEG	Fibromyalgia	Nociplastic	38	n.rep.	n.rep.	Higher beta power for patints (dorsal anterior cingulate cortex) compared with controls Lower beta lagged phase coherence and granger causality for patients compared with controls Lower pain after tDCS (stronger reduction for active compared with sham tDCS) Higher beta power after tDCS in patients compared with controls (anterior cingulate cortex) Higher beta and gamma connectivity after tDCS in patients compared with controls (dorsal anterior cingulate cortex and pregenual anterior cingulate cortex)
De Vries 2013 21	Cross-sectional observational	—	EEG	Chronic pancreatitis	Mixed	32	64.8	n.rep.	Lower PAF in patients compared with controls Correlation between pain duration and PAF
Di Pietro et al. 2018 22	Cross-sectional observational	—	EEG	Trigeminal neuropathy	Neuropathic	40	66	3.7	Higher power in patients (4-25 Hz) compared with controls, most marked in the theta and low alpha range Correlation between pain intensity and alpha (T4) and beta power (FC6, CP6, O1) for selected electrodes
Fallon et al. 2018 23	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	37	115.2	n.rep.	Higher theta power in patients compared with controls Correlation between frontal theta power and measures of tenderness
Fauchon et al. 2022 24	Cross-sectional observational	—	MEG	Neuropathic pain of various etiologies	Neuropathic	100	161.4	6.1	Higher alpha power in patients compared with controls Lower PAF in male patients compared with male controls Higher PAF for women compared with men with neuropathic pain Negative correlation between PAF and pain intensity in the neuropathic pain group (posterior insula; right TPJ)
Feng et al. 2021 25	Descriptive	—	EEG	Low back pain	Mixed	27	67.2	n.rep.	Negative correlation between central alpha power and pain intensity
Ferdek et al. 2019 26	Cross-sectional observational	—	EEG	Endometriosis	Mixed	37	88.8	6.8	Higher beta connectivity in patients compared with controls (left dorsolateral prefrontal cortex—left somatosensory cortex; left somatosensory cortex—orbitofrontal cortex and right temporal cortex)
Freye and Levy 2006 27	Longitudinal descriptive	Tramadol	EEG	Osteoarthritis	Mixed	19	n.rep.	7.8	Higher alpha and beta power after tramadol administration Lower pain after tramadol
González-Roldán et al. 2016 30	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	38	225.6	6.0	Higher beta power in patients compared with controls Higher theta and beta connectivity (left hemisphere) in patients compared with controls
González-Villar 2020 31	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	94	n.rep.	7.0	Higher beta connectivity in patients compared with controls Microstate analysis: shorter microstate 1 occurrence and coverage in patients compared with controls
Gram 2017 32	Longitudinal descriptive	Surgery	EEG	Chronic hip pain	Mixed	81	n.rep.	n.rep.	No post-OP power or connectivity differences between opioid treatment responders and nonresponders Machine learning: No distinction of responders and nonresponders based on EEG features
Graversen 2012 33	Longitudinal descriptive	Pregabalin vs placebo	EEG	Chronic pancreatitis	Mixed	28	116	3.8	Higher theta power after pregabalin (not after placebo) Lower pain for most patients (individual level data only)
Hargrove 2010 35	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	170	132	n.rep.	Higher beta power in patients compared with controls
Heitmann 2022 37	Longitudinal descriptive	Interdisciplinary multimodal pain therapy	EEG	Mixed	Mixed	41	101.6	5.4	Lower pain after intervention No correlation between pain changes and PAF or power changes Network analysis: correlation between pain changes and increase in theta global network efficiency
Hsiao 2017 38	Cross-sectional observational	—	MEG	Fibromyalgia	Nociplastic	56	87.6	5.3	Lower theta connectivity (insula—DMN bilaterally) for patients compared with controls No correlation between pain intensity and duration and connectivity measures Correlation between number of tender points and connectivity (right insula—DMN)
Hunter 2009 39	Longitudinal descriptive	Duloxetine vs placebo	EEG	Fibromyalgia	Nociplastic	12	n.rep.	6.7	Regression analysis: prediction of pain improvement by theta-cordance change in duloxetine group after 12 wk (also after controlling for depression)
Iwatsuki 2021 40	Cross-sectional observational	—	MEG	CPRS	Nociplastic	42	n.rep.	n.rep.	Correlation between pain intensity and alpha/theta connectivity (alpha: right SII; precuneus/insula, theta: right SII; posterior cingulate cortex)
Jensen 2013 41	Longitudinal descriptive	NFB (different protocols)	EEG	Spinal cord injury	Neuropathic	10	n.rep.	5.9	Lower theta and higher alpha power after NFB Lower pain after NFB
Jensen 2021 43	(Randomized) controlled trial	Different cognitive and/or behavioral interventions	EEG	Mixed	Mixed	147	n.rep.	4.7	Lower theta power after intervention only for pain education group Lower pain intensity across all interventions No correlation between changes in pain intensity and EEG measures
Jensen 2013 42	Cross-sectional observational	—	EEG	Spinal cord injury	Mixed	82	n.rep.	n.rep.	Lower alpha power for patients compared with controls, no differences in other frequency bands No correlations between pain intensity and global power measures Correlation between pain intensity and frontal alpha power
Juel 2017 44	(Randomized) controlled trial	Acupuncture vs sham stimulations	EEG	Chronic pancreatitis	Mixed	15	141.6	n.rep.	No changes in theta, alpha, beta power after intervention Lower pain after intervention
Kayiran 2010 45	Longitudinal descriptive	NFB vs escitalopram	EEG	Fibromyalgia	Nociplastic	54	55.3	8.9	Lower theta/SMR ratio after NFB No changes in theta, alpha or beta power
Kim 2020 47	Cross-sectional observational	—	MEG	MS with chronic pain	Mixed	63	n.rep.	3.7	Abnormalities in theta, alpha, beta, and gamma cross-network functional coupling for patients compared with controls
Kim 2019 46	Cross-sectional observational	—	MEG	MS with chronic pain	Mixed	53	n.rep.	3.4	Higher alpha and lower beta power in patients compared with controls (most prominently in thalamus, insula, and right TPJ) Higher alpha power in neuropathic compared with nonneuropathic pain patients Correlation between pain intensity and alpha power
Kim 2021 48	Longitudinal descriptive	Non-invasive painless signaling therapy	EEG	Failed back surgery syndrome	Mixed	11	n.rep.	4.3	Higher alpha and beta power after treatment in responder group (pain reduction)
Kisler 2020 49	Cross-sectional observational	—	MEG	Ankylosing spondylitis	Mixed	76	164.4	3.6	Higher theta and lower gamma in patients compared with controls Higher alpha power in subgroup of neuropathic pain patients compared with other patients and to controls Correlation between pain intensity and alpha power in mixed-pain patients
Klug 2011 51	Cross-sectional observational	—	EEG	Somatoform pain disorder	Nociplastic	30	n.rep.	n.rep.	Lower beta power in patients compared with controls
Lee 2018 56	Descriptive	—	EEG	Fibromyalgia	Nociplastic	10	n.rep.	n.rep.	Network analysis: “Explosive synchronization”—characteristics found in patients, correlation between pain intensity and these characteristics
Levitt 2020 59	Cross-sectional observational	—	EEG	Chronic lumbar radiculopathy	Neuropathic	37	n.rep.	7.1	Gamma-band differences between patients and controls for number, power, and frequency span of transient spectral events No differences between patients and controls for power spectral density, coherence, or phase-amplitude coupling Machine learning: binary SVM (patients vs controls) and 3-way-classifier (lumbar pain vs radiculopathy vs controls) based on power spectral density, coherence, and phase-amplitude coupling performed significantly better than chance
Lim 2016 60	Cross-sectional observational	—	MEG	Fibromyalgia	Nociplastic	36	36.6	n.rep.	Lower PAF in patients compared with controls Higher theta, beta, gamma power in patients compared with controls (no test statistics provided)
Lopez 2015 63	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	26	n.rep.	n.rep.	Lower theta and alpha in patients compared with controls Higher beta in patients compared with controls
Martín-Brufau 2021 66	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	46	n.rep.	n.rep.	Lower theta and alpha in patients compared with controls Higher beta in patients compared with controls
May 2021 67	Cross-sectional observational	—	EEG	Mixed	Mixed	185	121.9	5.7	Microstate analysis: less dominant role of microstate D in patients (lower time coverage, lower frequency of occurrence, lower global explained variance, shorter mean duration); lower transition probability from microstates A, B, and E to microstate D in patients Microstate analysis: no correlation between microstate D measures and clinical characteristics
Meneses 2016 68	Cross-sectional observational	—	EEG	Rheumatoid arthritis	Mixed	42	107.4	n.rep.	Higher alpha and theta power in patients compared with controls No correlations between power and pain measures
Michels 2011 69	Cross-sectional observational, longitudinal descriptive	Thalamotomy	EEG	Neuropathic pain of various etiologies	Neuropathic	38	55.2	6.5	Higher theta and alpha power in all patients compared with controls Higher beta and gamma power in patients with low pain relief (after surgery) compared with controls Correlation between theta and beta power with pain intensity presurgery normalization of power in all frequency bands postsurgery in patients with high pain relief
Nomura 1999 74	Cross-sectional observational	—	EEG	Inflammatory bowel syndrome	Nociplastic	48	n.rep.	n.rep.	Lower alpha and higher beta in patients compared with controls
Oga 2002 75	Longitudinal descriptive	Ketamine i.v.	EEG	Neuropathic pain of various etiologies	Neuropathic	10	n.rep.	n.rep.	Lower absolute theta power and higher relative theta power after ketamine Lower absolute/relative alpha power and higher absolute/relative beta power after ketamine Pain relief after ketamine Correlation between changes in pain intensity and alpha power (right central electrode)
Olesen 2011 76	Cross-sectional observational	—	EEG	Chronic pancreatitis	Mixed	46	120	3.8	Higher theta and alpha power in patients compared with controls Lower beta power in patients compared with controls
Parker 2021 79	Cross-sectional observational	—	MEG	Neuropathic pain of various etiologies	Neuropathic	13	n.rep.	n.rep.	Higher PAF after DRGS Higher theta power (DRGS-OFF) in patients with severe pain compared with mild pain Lower alpha power (DRGS-OFF) in patients with severe pain compared with moderate pain
Parker 2021 80	(Randomized) controlled trial	DRGS + (tDCS vs sham)	EEG	Neuropathic pain of various etiologies	Neuropathic	16	n.rep.	6.0	Higher beta power after DRGS-ON vs OFF and after DRGS-ON/active tDCS vs DRGS-ON/sham stimulation Lower pain across all interventions (highest reduction for DRGS-ON paired with active tDCS)
Patel 2020 82	Longitudinal “comparative”	NFB	EEG	Mixed	Mixed	8	n.rep.	n.rep.	No differences in alpha power between patients and controls (baseline and longitudinally) Alpha power fluctuation analysis: no differences in fractional occupancy, dwell time, distribution and transition probability between patients and controls (baseline and longitudinally) Network analysis: negative correlations of changes in pain with fractional occupancy, mean dwell times, and transition probability from low to high alpha state
Prichep 2011 89	Longitudinal descriptive	Various treatments for neuropathic pain	EEG	Neuropathic pain of various etiologies	Neuropathic	5	13.6	6.4	Higher alpha power after intervention
Prichep 2018 90	Cross-sectional observational	—	EEG	Mixed	Mixed	154	100.8	5.0	Higher theta and low alpha power in patients compared with controls Higher theta connectivity in patients compared with controls
Prinsloo 2017 91	(Randomized) controlled trial	NFB vs waiting list	EEG	Chemotherapy-induced peripheral neuropathy	Neuropathic	62	n.rep.	4.7	Higher alpha power after NFB Lower beta power after NFB Greater decrease in worst pain in NFB group, but similar outcomes for groups in average pain
Santana 2021 95	Cross-sectional observational	—	EEG	Sickle cell disease with chronic hip pain	Mixed	40	55.5	5.5	Network analysis: lower full synchronization time in patients compared with controls
Sarnthein 2006 96	Cross-sectional observational, longitudinal descriptive	Thalamotomy	EEG	Neuropathic pain of various etiologies	Neuropathic	30	n.rep.	n.rep.	Lower PAF and theta power in patients compared with controls Lower theta power and lower pain 12 mo after surgery
Schmidt 2015 98	Longitudinal descriptive	Cognitive and/or behavioral intervention	EEG	Back pain	Mixed	21	n.rep.	5.9	No EEG differences after intervention Lower pain after intervention No correlation between EEG and pain
Schmidt 2012 97	Cross-sectional observational	—	EEG	Back pain	Mixed	74	n.rep.	n.rep.	No differences in PAF and overall power between patients and controls Correlation between overall power and pain intensity, no correlation with PAF
Stern 2006 100	Cross-sectional observational, longitudinal descriptive	Thalamotomy	EEG	Neuropathic pain of various etiologies	Neuropathic	32	n.rep.	7.1	Higher theta, alpha, beta power in patients compared with controls “Deoveractivation” (lowering of alpha and beta power) for subgroup of patients with high pain relief after surgery
Sufianov 2014 101	Cross-sectional observational, longitudinal descriptive	Epidural spinal cord stimulation	EEG	Failed back surgery syndrome	Mixed	40	38.4	7.2	Lower PAF and theta power in patients compared with controls Higher beta power in patients compared with controls Lower theta power and higher PAF after intervention Lower pain after intervention
Ta Dinh 2019 102	Cross-sectional observational	—	EEG	Mixed	Mixed	185	121.9	5.7	No difference in power and PAF between patients and controls Higher theta and gamma connectivity in patients compared with controls Network analysis: decrease in global efficiency in gamma frequencies in patients compared with controls
Teixeira 2021 103	Cross-sectional observational	—	EEG	Neuropathic pain of various etiologies	Neuropathic	22	n.rep.	4.2	Lower beta power in patients compared with controls Negative correlation between beta power and pain intensity in subgroup of patients with more severe pain
Teixera 2022 104	Descriptive	—	EEG	Low back pain	Mixed	30	n.rep.	4.7	Regression analysis: prediction of pain intensity by theta power (while controlling for depression)
Thibaut 2017 105	Cross-sectional observational	tPCS + tDCS vs tPCS vs tDCS vs sham	EEG	Chronic pancreatitis	Mixed	54	n.rep.	n.rep.	Higher theta, alpha, beta power in patients compared with controls Lower alpha/beta ratio in patients compared with controls No pain reduction after intervention
Uygur-Kucukseymen 2020 109	Descriptive	—	EEG	Fibromyalgia	Nociplastic	21	n.rep.	5.9	Negative correlation between pain intensity and alpha and beta power
Van den Broeke 2013 110	Cross-sectional observational	—	EEG	Chronic pain after breast cancer treatment	Mixed	19	n.rep.	5.0	Higher alpha power in patients compared with controls No differences in PAF No correlation between pain intensity and alpha power
Vanneste 2021 112	Cross-sectional observational	—	EEG	Neuropathic pain of various etiologies	Neuropathic	100	n.rep.	5.8	Higher theta, alpha, beta, gamma power in patients compared with controls Correlations between pain intensity and gamma and beta power (not for theta and alpha) Higher theta and lower alpha functional connectivity in patients compared with controls Source analysis: imbalance between ROIs involved in pain input and pain suppression
Vanneste 2017 113	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	88	14.9	n.rep.	Lower alpha power in patients compared with controls Higher beta power in patients compared with controls Differences in alpha connectivity patterns in patients compared with controls for several regions (higher/lower depending on ROI)
Vanneste 2018 114	Cross-sectional observational	—	EEG	Neuropathic pain of various etiologies	Neuropathic	342	n.rep.	n.rep.	Higher theta, beta, and gamma power in patients compared with controls (different ROIs) Lower alpha power in patients compared with controls Cross-frequency coupling: increased nesting of theta–beta and theta–gamma in patients compared with controls Machine learning (SVM): discrimination of patients and controls based on EEG features
Villafaina 2019 115	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	62	n.rep.	5.9	Lower alpha power in patients compared with controls Negative correlation between pain scores and alpha power (single electrode level)
Witjes 2021 120	Cross-sectional observational	—	MEG	Failed back surgery syndrome	Mixed	46	120	6.1	Higher alpha power ratio (low/high alpha) in patients compared with controls No differences in PAF or theta, alpha, beta, gamma power between patients and controls No correlation between pain intensity or duration and alpha power ratio
Yüksel 2019 122	Cross-sectional observational, longitudinal descriptive	tENS vs acupuncture	EEG	Fibromyalgia	Nociplastic	63	4,5	5.0	Lower theta power in both patient groups compared with controls Lower alpha and beta power only in acupuncture group compared with controls Lower pain across interventions Correlation between theta power and alpha power with pain in acupuncture group Higher alpha power across interventions, diverging effects for alpha power in TENS group (higher/lower depending on ROI), higher beta power in acupuncture group
Zhou 2018 124	Cross-sectional observational	—	EEG	Postherpetic neuralgia	Neuropathic	28	11.9	5.0	Higher gamma power in patients compared with controls (not significant after controlling for anxiety) Correlation between pain intensity and gamma power
Zolezzi 2021 125	Descriptive	—	EEG	Neuropathic pain of various etiologies	Neuropathic	35	n.rep.	n.rep.	Machine learning (SVM): classification into 3 groups (low, moderate, high pain) with high accuracy based on linear (frequency-specific power) and nonlinear (“approximate entropy”) features; frontal beta power most relevant feature
Zortea 2021 126	Cross-sectional observational	—	EEG	Fibromyalgia	Nociplastic	47	n.rep.	8.8	No differences for theta, alpha, or beta power between opioid users compared with nonopioid users Lower theta and beta peak amplitudes in opioid users compared with nonopioid users No differences in pain between the groups

DRGS, dorsal root ganglion stimulation; EC, eyes closed; EO, eyes open; MS, multiple sclerosis; n.rep, not reported; NFB, neurofeedback; PAF, peak alpha frequency; ROI, region of interest; SMR, sensorimotor; SVM, support vector machine; tACS, transcranial alternating current stimulation; tDCS, transcranial direct current stimulation; TENS, transcutaneous electrical nerve stimulation; TPJ, temporoparietal junction; tRNS, transcranial random noise stimulation.

3.3. Risk of bias and open science practices

The results of the risk of bias assessment are summarized in Figure 3. Scores of individual studies are presented in Table S1, https://links.lww.com/PAIN/B751.

Figure 3.: Risk of bias of included studies. M/EEG, magneto-/electroencephalography.

In the domain “selection of study participants,” the most significant risk of bias arose from “case representativeness” because most studies provided no description of their sampling/recruiting strategy. For the item “case definition,” risk of bias was low for most studies because pain diagnoses and diagnostic criteria were specified frequently. For the items “matching of controls” and “definition of controls,” mixed results were obtained. In the domain “comparability/confounders,” risk of bias was high for most studies because frequent comorbidities of chronic pain (eg, depression or anxiety) were rarely controlled for. Regarding “outcome data,” risk of bias was “low” for the items “data acquisition/processing” and “appropriate statistical test” for most studies. However, most studies had a high risk of bias regarding the “assessment of outcome” because preprocessing/analysis of M/EEG data was performed manually and blinding to condition, or clinical data were not explicitly stated.

Figure 4 shows the results of the evaluation of open science practices. Only few studies were preregistered, provided sample size calculations/justifications or offered data or code sharing. Almost half of the included studies adjusted P values to correct for multiple comparisons. However, in some studies, correction for multiple comparisons was performed for selected analyses only.

Figure 4.: Open science practices of included studies. The charts show absolute numbers of studies.

3.4. Data synthesis

For semiquantitative analysis of studies, we used modified albatross plots. Figure 5 explains the interpretation of albatross plots based on fictional data. A few studies reported diverging directions of effects for different regions of interest.^112,114 These studies were not included in the modified albatross plots, but their results are listed in Table 3. For studies using less common analysis techniques and outcome measures (eg, microstate analysis or machine learning algorithms) and for predictive biomarker studies, narrative data synthesis was performed due to the low overall number of studies and high heterogeneity of methods and outcome measures.

Figure 5.: Explanatory albatross plot based on fictional data. P values on the x-axis are displayed on a logarithmic scale (log10). n.s., not significant.

3.4.1. Semiquantitative data synthesis

3.4.1.1. Cross-sectional studies

Forty-four cross-sectional studies, which can serve the development of diagnostic biomarkers of chronic pain, evaluated differences in PAF, frequency-specific power, and/or connectivity between patients with chronic pain and healthy participants. Sample sizes varied between 13 and 342 participants (median, 44.0).

For PAF, 7 studies (58%) found no statistically significant differences between patients with chronic pain and healthy participants; 4 studies (33%) found lower and one study (8%) found higher peak frequencies in patients (Fig. 6).

Figure 6.: Peak alpha frequency differences between patients and healthy participants in cross-sectional studies. P values on the x-axis are displayed on a logarithmic scale (log10). n.s., not significant.

In Figure 7, cross-sectional results for frequency-specific power and connectivity are depicted. In 14 studies (47%), theta power was higher for patients compared with healthy participants, closely followed by nonsignificant results (n = 13; 43%). Only 3 studies (10%) found lower theta power for patients. For alpha power, no clear differences were observed (n = 13; 39% no difference, higher n = 11; 33%, lower n = 9; 27%). Regarding beta power, most studies found higher values for patients (n = 14; 47%), whereas 11 studies (37%) found no differences and 5 (17%) studies found lower values for patients compared with healthy participants. For gamma power, 9 studies (60%) found no group differences, 5 studies (33%) reported higher gamma power, and 1 (7%) study reported lower gamma in patients than in healthy participants.

Figure 7.: Power and connectivity differences between patients and healthy participants in cross-sectional studies. P values on the x-axis are displayed on a logarithmic scale (log10). n.s., not significant.

Compared with frequency-specific power, the number of cross-sectional studies analyzing connectivity differences between patients and healthy participants was lower. No clear differences across studies could be observed for any frequency band. However, more studies reported higher than lower connectivity at theta frequencies for patients (theta: n = 4; 40% higher, n = 4; 40% no difference, n = 2; 20% lower/alpha: n = 8; 80% no difference, n = 2; 20% lower, n = 0 higher/beta: 6; 55% no difference, n = 3; 27% higher, n = 2; 18% lower/gamma: n = 7; 78% no difference, n = 1; 11% higher, n = 1; 11% lower).

In summary, more cross-sectional studies reported higher theta and beta power in patients with chronic pain compared with lower power or nonsignificant results. Although most studies reported nonsignificant results for the remaining parameters, more studies found higher than lower gamma power and theta connectivity and lower than higher peak alpha frequency in patients with chronic pain.

3.4.1.2. Longitudinal studies

Twenty-two (interventional) longitudinal studies, which can aid the development of monitoring and predictive biomarkers, reported pain relief and premeasurements and postmeasurements of frequency-specific power (Fig. 8). No study analyzed connectivity longitudinally. Sample size (patients only) varied between 6 and 147 patients (median, 24.5).

Figure 8.: Power changes with pain relief in longitudinal studies. On the right (“[higher]”) part of each albatross plot, studies with postinterventional power increases are displayed. On the left ([“lower”]) part of each plot, studies with postinterventional power decreases are displayed. Only studies that found a postinterventional pain relief were included. P values on the x-axis are displayed on a logarithmic scale (log10). n.s., not significant.

No trend toward higher or lower power values after intervention could be observed across studies for any frequency band (theta: n = 7; 43.8% no difference, n = 6; 37.5% lower, n = 3; 18.8% higher/alpha: n = 5; 31.3% no difference, n = 5; 31.3% lower, n = 6; 37.5% higher/beta: n = 7; 43.8% no difference, n = 4; 25% lower, n = 5; 31.3% higher/gamma: n = 1; 33.3% no difference, n = 2; 66.7% lower, n = 0 higher).

In summary, included studies did not show consistent longitudinal power changes in patients undergoing interventions. However, more studies reported decreases than increases in theta power.

3.4.1.3. Descriptive studies

Twenty-five descriptive studies, which can aid the development of diagnostic and/or monitoring biomarkers correlated M/EEG measures with pain intensity (Fig. 9). Sample sizes ranged from 12 to 103 patients (median 27.0). For all frequency bands, most studies found nonsignificant correlations with pain intensity, whereas for the alpha band, more studies reported a negative correlation than a positive correlation (theta: n = 10; 91% no correlation, n = 1; 9% positive correlation, n = 0 negative correlation/alpha: n = 9; 50% no correlation, n = 7; 38.9% negative correlation, n = 2; 11.1% positive correlation/beta: n = 9; 75% no correlation, n = 3; 25% positive correlation, n = 0 negative correlation/gamma: n = 3; 60% no correlation, n = 2; 40% positive correlation, n = 0 negative correlation). Similarly, for connectivity, mainly nonsignificant correlations were observed for all frequency bands (theta: n = 4; 80% no correlation, n = 1; 20% negative correlation, n = 0 positive correlation/alpha: n = 2; 50% no correlation, n = 1; 25% negative correlation, n = 1; 25% positive correlation/beta: n = 4; 66.7% no correlation, n = 2; 33.3% negative correlation, n = 0 positive correlation/gamma: n = 3; 75% no correlation, n = 1; 25% negative correlation, n = 0 positive correlation).

Figure 9.: Correlations of power and connectivity with pain intensity in descriptive studies. On the right (“[pos.]”) part of each albatross plot, studies with positive correlations of pain intensity and power/connectivity are displayed. On the left ([“neg.”]), part of each plot, studies with negative correlations of pain intensity and power/connectivity are displayed. P values on the x-axis are displayed on a logarithmic scale (log10). n.s., not significant.

In summary, neither for power nor for connectivity measures, a clear relation with pain intensity was found across studies. However, more negative than positive correlations of pain intensity and alpha power were observed.

3.4.2. Narrative synthesis of studies using microstate analysis, network analysis, and machine learning approaches

In recent years, a variety of advanced analysis methods for M/EEG data have been applied to assess brain function in chronic pain populations. Furthermore, machine learning approaches have been used to differentiate between patients and healthy participants based on M/EEG data. These studies include cross-sectional, longitudinal, and descriptive studies, which can yield diagnostic, monitoring, and predictive biomarkers of chronic pain.

Two cross-sectional studies performed microstate analysis of EEG recordings. Although traditional quantitative resting-state EEG analysis focuses on static measures of brain activity over minutes, it is increasingly recognized that brain activity varies over short time intervals and that corresponding temporal dynamics provide physiologically relevant information. It has been shown that these dynamics can be characterized as sequences of recurrent topographies of EEG activity termed microstates, which are believed to reflect large-scale brain network activity. Microstate analysis describes EEG activity as a sequence of a limited number of microstates that remain stable for tens of milliseconds before abruptly transitioning to another microstate. Electroencephalography resting-state activity is usually described well as a series of 4 to 6 microstates, which are remarkably similar across participants and labeled with letters. One microstate study detected lower occurrence and time coverage of microstate C in patients with chronic widespread pain compared with healthy participants.³¹ Another study found a less predominant role of microstate D in patients with chronic back pain than in healthy participants.⁶⁷ In the field of network analysis (an overarching term for analysis of global measures of brain connectivity mainly based on graph theory), one cross-sectional study found a decrease in global efficiency in gamma frequencies in patients with chronic pain compared with healthy participants.¹⁰² Another cross-sectional study observed a lower “full synchronization time” (described as the time required for a network to become approximately complete) in sickle cell patients with chronic hip pain.⁹⁵ In a longitudinal study examining changes in brain activity following multimodal pain therapy, a correlation between pain decrease and theta global efficiency increase was observed.³⁷ A longitudinal neurofeedback study found negative correlations of changes in pain with fractional occupancy, mean dwell times, and transition probability from low to high alpha state.⁸² A descriptive study described an association of “explosive synchronization” characteristics and pain intensity in patients with fibromyalgia.⁵⁶

M/EEG studies using machine learning algorithms have addressed diagnostic and predictive biomarker functions. Cross-sectional studies have been shown to differentiate better than chance between chronic pain populations and healthy participants in several studies (¹¹⁴accuracy 92.5%¹⁰²; accuracy 57%⁵⁹; accuracy 82.3%). Two studies applied machine learning to differentiate between different patient groups. In one of these studies, a 3-way classifier distinguished between patients with lumbar pain, patients with radiculopathy, and healthy participants with a cross-validation accuracy of 71.9%.⁵⁹ The other study differentiated with high accuracy between patients with low (97%), moderate (96%), and high (96%) pain.¹²⁵ Only 2 studies used EEG as predictive biomarkers examining response to treatment. One study in patients with lower back pain found no predictive value of baseline EEG regarding pain changes after cognitive/behavioral interventions.¹⁵ In another study, preoperative EEG characteristics could not predict pain-related outcome after surgery in patients with chronic hip pain.³²

4. Discussion

4.1. Main findings

In this study, we summarized evidence on resting-state M/EEG as biomarkers of chronic pain. To this end, we conducted a systematic review of cross-sectional, longitudinal, and descriptive studies of patients with chronic pain using resting-state M/EEG as a measure of brain function. Overall, this approach revealed differences between study and, thus, potential biomarker types. For cross-sectional studies that can serve to develop diagnostic biomarkers, we found increased theta and beta power in patients with chronic pain. For longitudinal studies that can yield monitoring and/or predictive biomarkers, we found no clear associations of pain relief with M/EEG measures. Similarly, descriptive studies that can yield diagnostic or monitoring biomarkers showed no clear correlations of pain intensity with M/EEG measures. Notable findings regarding frequently discussed biomarker candidates in pain research included some evidence for lower peak alpha frequency, higher gamma power, and higher theta connectivity in patients with chronic pain in cross-sectional studies. Furthermore, we found some evidence for lower theta power along with pain relief in longitudinal studies and for a negative correlation of alpha power with pain intensity. However, for those parameters, studies reported mainly nonsignificant results.

A few previous systematic reviews of M/EEG findings in chronic pain have been published. However, their scope differed from this systematic review because of the focus on certain pain types and/or the joint examination of different biomarker types. One systematic review focused on patients with neuropathic pain caused by spinal cord injury and found evidence suggesting a reduction of PAF.⁸¹ Another review⁸⁷ summarized resting-state and event-related EEG studies in a variety of patients with chronic pain. Regarding resting-state studies, they found evidence for an increase of theta and alpha power in patients. Recently, a third review focused on resting-state EEG studies in patients with neuropathic pain. They found power increases at theta and—to a lesser degree—at high beta frequencies. In addition, they observed power decreases at high alpha and low beta frequencies.⁷² No clear correlations with pain intensity were found.

In general, these results are in line with our findings. The most consistent finding across reviews was an increase in theta power in patients. In addition, we found increased beta power and some evidence for increases of gamma power and theta connectivity and toward a slowing of the PAF. These additional findings of this review might be the result of significantly larger number of studies included in this review (76 studies) than in the previous reviews (4-15 studies). Moreover, the semiquantitative approach of this review might be more sensitive than qualitative methods.

Our central finding of increased theta and beta power is compatible with the thalamocortical dysrhythmia concept of neuropsychiatric disorders,^61,62 which is supported by animal studies in chronic pain models.^55,93 The concept proposes that abnormal thalamocortical theta activity yields abnormal oscillations at beta/gamma frequencies, which eventually result in different neuropsychiatric symptoms, including ongoing pain. However, the concept has been proposed to account not only for chronic pain but also for various neuropsychiatric disorders. Moreover, opioid and antiepileptic treatment, frequently used in patients with chronic pain, can also yield increases of theta activity.^65,94 In addition, the included studies in our review reported no consistent correlation between theta or beta power and pain intensity, and pain relief was not associated with power changes in those frequency bands. Thus, the specificity of increased theta and beta power for chronic pain remains unclear.

Besides frequency-specific power and connectivity, PAF has been discussed as a biomarker of chronic^21,96 and experimental pain.²⁸ Across the studies included in our review, we found some evidence for a slowing of PAF in patients, which can also be interpreted in the context of the thalamocortical dysrhythmia concept. However, peak alpha frequency has also been related to treatment response in attention deficit hyperactivity disorder² depression,³ cognitive functioning,⁵⁰ and age.⁴ Hence, also the specificity of PAF slowing for chronic pain remains unclear.

4.2. Risk of bias and limitations

Risk of bias assessment revealed limitations of many studies included in this review. First, most studies have small sample sizes (median 38.0) and a priori sample size calculations were rarely reported. Correction for multiple comparisons was performed in only about half of studies. Second, in almost all studies, M/EEG preprocessing and analysis was performed manually and was thus not well standardized. Third, only few studies accounted for potential confounds by medication effects, demographic factors (eg, age or sex/gender¹⁰⁶), and neuropsychiatric comorbidities.

Limitations also apply to this systematic review itself. First, most studies included patients with different types and etiologies of chronic pain. Although such mixed pain types and populations reflect clinical reality, we did not take possible differences in brain mechanisms of different pain etiologies or types (eg, neuropathic pain and nociplastic pain) into account. However, color coding of studies with different pain types in the albatross plots did not indicate clear differences in results between pain types. Second, the designs of longitudinal studies were highly heterogeneous. The types of interventions and the timing of postinterventional assessments varied considerably. This heterogeneity might have obscured effects of single treatment methods. Third, for semiquantitative data synthesis by means of albatross plots, we summarized studies regardless of whether they reported global or local M/EEG measures. This might obscure contributions of specific brain regions and/or networks to the neural representation of chronic pain. Fourth, we excluded studies of patients with chronic pain caused by severe neurological comorbidity (eg, severe spinal cord injury) to increase specificity of results. Fifth, a few studies (not represented in the albatross plots) reported diverging effects for M/EEG measures depending on the region of interest. However, graphical or numerical inclusion of these results in our analysis would have led to misrepresentation of results being interpreted as complex network changes in these studies. Sixth, meta-analysis would have further increased the validity of our findings. However, because of reported outcome measures (eg, for cross-sectional comparisons of theta power, 47% of all studies but 79% of studies with negative results did not report the minimally required parameters for meta-analysis) and heterogeneous study designs, we could not perform formal meta-analysis.

4.3. Outlook and recommendations

This systematic review might help to guide future M/EEG studies on the development of pain biomarkers.

4.3.1. Differentiating biomarker types

Cross-sectional studies indicated increased theta and beta power in patients with chronic pain, which might serve the development of diagnostic biomarkers. By contrast, longitudinal studies that aim at monitoring rather than diagnostic biomarkers did not show a clear association between changes in pain intensity and changes of brain activity at any frequency. This dissociation suggests that different types of biomarkers might be served by different neuronal phenomena. Hence, different biomarkers should be considered for different functions. Moreover, most studies aimed at diagnostic and monitoring biomarkers. However, other biomarker types, such as predictive or prognostic biomarkers, might be at least equally important in clinical practice. Thus, we might broaden the scope of M/EEG biomarker development toward these biomarker types using longitudinal study designs with baseline or repeated M/EEG measurements. Recent large-scale studies in depression, in which resting-state EEG data predicted response to antidepressant treatment,^121,123 could serve as a blueprint for pain research here.

4.3.2. Assessing biomarker specificity

The risk of bias assessment indicated that most included studies did not control for potential neuropsychiatric confounders like depression or anxiety. Thus, the specificity of the observed findings remains unclear. However, specificity is a key criterion in biomarker development.¹¹¹ Consequently, findings in patients with chronic pain should be related to findings in other neuropsychiatric conditions, especially as chronic pain and affective disorders frequently co-occur.¹³ A recent review on resting-state EEG activity in neuropsychiatric disorders found increases in delta and theta power in depression, schizophrenia, attention-deficit hyperactivity disorder, and obsessive–compulsive disorder.⁷³ An increase of theta power was also the most consistent finding in our review. Hence, large-scale studies across neuropsychiatric disorders, including chronic pain, are needed to carve out the specific M/EEG signatures of chronic pain. However, it is important to note that for clinical utility, specificity is not mandatory. Clinicians are very much used to make decisions based on patterns of unspecific information.

4.3.3. Investigating connectivity biomarkers

Most studies in this review focused on oscillatory brain activity in chronic pain. Fewer studies assessed connectivity, and we did not find a clear connectivity pattern that could serve as a biomarker of chronic pain. This might at least partly be the result of the complexity and methodological variability of connectivity analyses. However, the complexity of chronic pain and accumulating evidence from neuroimaging studies indicate that changes of connectivity in large-scale brain networks figure prominently in the pathology of chronic pain.^52,58,88 For instance, a recent functional magnetic resonance imaging study on long-lasting experimental and clinical pain⁵⁷ provided convincing evidence for a functional connectivity signature of ongoing pain. Thus, connectivity and network measures are particularly promising for the development of biomarkers of chronic pain.

4.3.4. Enhancing biomarker reproducibility

The sample sizes of studies included in this systematic review is rather small (median n = 38). Moreover, only very few studies have implemented most recent open science measures to enhance the reproducibility and transparency of the findings.⁷¹ However, timely approaches for the development of chronic pain biomarkers include machine learning, multivariate models, and multimodal/composite biomarker studies. These approaches essentially depend on large data sets, standardized clinical assessment, and fully transparent and standardized neuroimaging methods.⁸⁵ In M/EEG biomarker development, this might include performance and reporting of studies according to consensus statements,⁸⁵ a standardized Data structure (EEG-BIDS),⁸⁴ automatic preprocessing and analysis pipelines,^29,83,86 and sharing of code and data.

4.3.5. Exploring composite biomarkers

We here reviewed M/EEG studies on brain function in chronic pain. However, chronic pain is not only associated with changes of brain function but also associated with brain structure,⁵⁴ sensory functions,¹¹⁸ psychological and social factors,¹³ and genetic information.⁷⁰ Considering these different biopsychosocial determinants of chronic pain, integrating them into multimodal composite biomarkers is an evident next step.¹⁰⁷ Such composite biomarkers, which integrate information from M/EEG recordings with other types of information, promise to enhance the sensitivity and specificity of biomarkers and might therefore be particularly worth exploring.

5. Conclusions

The development of biomarkers has been recognized as a crucial step toward optimizing the treatment of chronic pain.^14,107 With accumulating evidence for substantial changes of brain structure and function in chronic pain, noninvasive, brain-based measures can likely significantly contribute to the development of pain biomarkers. Because of its scalability and ease of use in clinical settings, M/EEG holds great potential as pain biomarker.^14,92,111 In this systematic review, we therefore summarized current evidence on resting-state M/EEG studies as biomarkers of chronic pain. In cross-sectional studies, we found evidence for increased theta and beta power as a diagnostic biomarker in patients with chronic pain. With respect to monitoring or predictive M/EEG biomarkers, evidence from longitudinal and descriptive studies is sparse and inconsistent so far. To proceed and better exploit the potential of M/EEG as biomarkers of pain, large-scale studies that differentiate biomarker types, assess biomarker specificity, investigate brain network connectivity, adhere to highest standards of transparency/reproducibility, and explore the potential of composite biomarkers are most promising.

Conflict of interest statement

The authors have no conflict of interest to declare.