Exploring electroencephalographic chronic pain biomarkers: a mega-analysis

Abstract

Background: Chronic pain is associated with alterations in brain function, offering promising avenues for advancing diagnostic and therapeutic strategies. In particular, these alterations may serve as brain-based biomarkers to support diagnosis, guide treatment decisions and monitor clinical courses of chronic pain.

Methods: Motivated by this potential, this study analysed associations between chronic pain and changes of large-scale brain network function using resting-state electroencephalography (EEG) from 614 individuals with chronic pain, collected by research groups from Australia, Germany, Israel, New Zealand, and the US.

Findings: Employing a discovery-replication approach, we found limited replicability of associations between pain intensity and brain network connectivity. However, a mega-analysis combining all datasets revealed robust associations between pain intensity and large-scale brain network connectivity at theta frequencies and including the limbic network. Additionally, multivariate analyses identified connectivity patterns spanning theta, alpha, and beta frequencies with strong evidence for associations with pain intensity. Variations and ablations of model features yielded deeper insights into the relative importance of distinct electrophysiological brain features in assessing chronic pain.

Interpretation: Our findings highlight challenges and provide guidance for developing EEG-based, scalable, and affordable biomarkers of chronic pain.

Funding: This project was funded by the Deutsche Forschungsgemeinschaft and the Technical University of Munich.

Keywords: Biomarkers; Chronic pain; Electroencephalography; Large-scale brain networks; Replicability.

Fig. 1

Research questions. We used EEG data from eight independent studies involving people with chronic pain to investigate how brain connectivity relates to pain intensity.

Fig. 2

Pain intensity and age distributions. For the distributions of pain intensity, it is additionally specified for each dataset to which time period prior to the assessment the ratings refer.

Fig. 3

Methods for estimating representative time series. (a) Conceptual diagram of signal mixing due to imperfect source reconstruction (top half) and unmixing procedure (bottom half). (b) Illustration of two variants of the proposed method. Variant 1: For each network pair, one corresponding pair of orthogonal representative time series is determined. Variant 2: For each individual network, one representative time series is determined which is orthogonal to multiple (orthogonal) time series representing activity in all other networks.

Fig. 4

Univariate correlations between standard EEG features and pain intensity. (a) Correlations between pain intensity and EEG features in the discovery set. Each tile's top number and colour represent the correlation coefficient, and the bottom number is the associated BF. (b) Correlations in the pooled replication sets. The meanings of numbers and colours match those in panel (a). (c) Correlations in the joint set. The meanings of numbers and colours match those of panel (a).

Fig. 5

Univariate correlations between pain intensity and brain network connectivity at theta, alpha, and beta frequencies. (a) Correlations between pain intensity and brain network connectivity in the discovery set. Each heatmap tile's top number and colour represent the correlation coefficient; the bottom number is the associated BF. (b) Correlations in the pooled replication sets. The meanings of numbers and colours match those of panel (a). (c) Correlations in the joint set. The meanings of numbers and colours match those of panel (a). SMN, somatomotor network; SN, salience network; FPN, frontoparietal network; DN, default network; LN, limbic network; DAN, dorsal attention network; VN, visual network.

Fig. 6

Associations between pain intensity and multivariate patterns of brain network connectivity. (a) In-sample, leave-one-participant-out cross-validated (LOO-CV) correlation between predicted and observed pain intensity in the discovery set. (b) Out-of-sample correlation between predicted and observed pain intensity in the pooled replication sets. (c) In-sample, LOO-CV correlation between predicted and observed pain intensity in the joint set. (d) Visualisation of corresponding model weights. The top number and colour of each tile represent the median of the weights across bootstrap samples. The bottom number represents the empirical p-value, i.e., the fraction of bootstrap samples for which the sign of this predictor differed from that of the median value. Only tiles with uncorrected empirical p < 0.05 are coloured.

Fig. 7

Associations between pain intensity and several multivariate patterns of brain network features. Each tile's top number and colour represent the leave-one-participant-out cross-validated correlation between predicted and observed pain intensity in the joint set. The top-centre tile shows the prediction-observation correlation for the model employing brain network connectivity features among the seven Yeo networks at theta, alpha, and beta frequencies as features. Bottom tiles show the prediction-observation correlations of models employing alternative network properties as features.