Projects

An EEG-based intrinsic brain network perspective on chronic pain: A multi-dataset study

Chronic pain is a multi-faceted and debilitating condition that imposes a large burden on patients and society. With a prevalence exceeding 20%, chronic pain is a major cause of disability and, as such, has a considerable socio-economic dimension (Breivik, Collett, Ventafridda, Cohen, & Gallacher, 2006; Kennedy, Roll, Schraudner, Murphy, & McPherson, 2014). However, the means for diagnosis and treatment of chronic pain remain limited. To address this deficit, we must deepen our understanding of the underlying pathology. Converging lines of evidence highlight the central role of the brain in the emergence of chronic pain (Kennedy et al., 2014; Ploner, Sorg, & Gross, 2017). However, the precise pathophysiological mechanisms that lead to chronic pain have yet to be identified. In this study, we intend to investigate the role of brain network function in chronic pain. Specifically, we want to use electroencephalography (EEG) to assess the function of intrinsic brain networks in chronic pain. Intrinsic brain networks are spatially distributed networks of connected brain areas that are synchronously activated during rest or particular tasks (Uddin, Yeo, & Spreng, 2019; Yeo et al., 2011). Abnormalities of the function of intrinsic brain networks have been observed in different neuropsychiatric disorders, including chronic pain (Menon, 2011). While these abnormalities have originally been identified using hemodynamic signals obtained from fMRI, we want to assess their involvement in chronic pain using EEG. EEG has a higher temporal resolution than fMRI and can, thus, specify the temporal and spectral characteristics of intrinsic brain network function. Moreover, since EEG is broadly available, potentially mobile, and cost-efficient, it can be easily scaled to large patient numbers in different settings. Here, we will adopt the most common definition of intrinsic brain networks as provided in (Yeo et al., 2011), which divides the brain into seven intrinsic brain networks. We will focus on four of these networks, which figure prominently in the pathology of neuropsychiatric disorders (Menon, 2011) and chronic pain (Brandl et al., 2022): the somatomotor (a.k.a. pericentral) network, the frontoparietal (a.k.a. lateral frontoparietal/ control/ central executive) network, the ventral attention (a.k.a. midcingulo-insular/ salience) network, and the default (a.k.a. medial frontoparietal) network. To analyze the function of these networks, we have developed a pipeline to evaluate EEG activity within and across networks. We will relate the function of the networks to the intensity of the pain experience. Moreover, we will relate brain network function to other dependent variables such as group (healthy vs. control) and depression. To assess the robustness of the effects, we will include multiple data sets from different sites. To this end, we obtained five additional data sets, which we will use to validate findings in our own data sets.

Assessing Chronic Pain in EEG using Pre-Trained Deep Neural Networks 

Identifying brain-based biomarkers of neuropsychiatric disorders is a key challenge in translational neuroscience. However, there is concern that brain-based biomarkers are context-dependent and do not generalize to independent data sets (Marek, Tervo-Clemmens et al. 2022).

We previously developed ML models predicting pain intensity from patterns of brain network connectivity (Bott, Zebhauser et al. 2024). While the replicability of a model trained on a dedicated discovery data set was limited, a model trained and tested on all available data provided strong evidence for a correlation between predicted and observed pain intensity. However, this model's explained variance was low.

The limited generalizability and low explained variance may reflect fundamental limitations of the information content of EEG signals about chronic pain intensity and/or suboptimal analytical choices. Here, we will address these issues. We will leverage state-of-the-art deep learning techniques to investigate how much information about chronic pain is contained in EEG signals. Unlike traditional methods that rely on pre-defined signal features such as band-specific power or connectivity, deep learning techniques use entire EEG signal segments as inputs. This approach minimizes the reliance on a priori assumptions, though it requires large amounts of data. However, only relatively few EEG data sets annotated with information about chronic pain are available.

To overcome this data limitation, we will pre-train the deep learning models on publicly available EEG data sets in a self-supervised manner, i.e., without incorporating information about chronic pain. In this way, the models will acquire "knowledge" about the general structure of EEG signals and transfer this knowledge to the task of predicting pain intensity in people with chronic pain. In addition, given our modeling approach, we will perform several analysis variants to understand which components of EEG signals are particularly informative about chronic pain. Specifically, we will assess how different configurations for processing EEG data before model training affect model performance. These configurations aim to assess the informative value of different frequency bands, spatial aggregations, and temporal scales.

Neurophysiological Mechanisms of Intra- and Inter-individual Variations of Pain 

The subjective experience of pain is inherently variable. It varies from person to person as well as within individuals. Moreover, intra-individual variations can manifest on different timescales, e.g., from moment to moment and from day to day. Studying variations of pain at the levels of person-to-person, day-to-day, and moment-to-moment, and the underlying brain mechanisms promise basic insights into how the human brain generates the subjective experience of pain. Moreover, it might help to understand and treat abnormal variations of pain perception in chronic pain disorders.

Previous studies investigating the brain representation of person-to-person and moment-to-moment variations yielded inconclusive findings, prompting a lively debate on this topic. In addition, the reliability of relationships between brain activity and pain variations remains largely untested. Moreover, intra-individual pain variations on longer time scales, e.g., from day to day, have not been investigated so far.

Here, we aim to fill these gaps and establish a robust and generalizable framework of the neurophysiological mechanisms encoding pain variations. We, therefore, comprehensively re-investigated the brain mechanisms of inter- and intra-individual variations of pain in 162 healthy participants (aged 18–80+). Experiments were performed twice, one month apart, allowing the assessment of intra-individual variations of pain on a longer time scale. Moreover, performing the experiments twice allowed us to assess the repeatability as a core element of the reliability of findings. In both experiments, participants received repeated brief painful laser stimuli, verbally rated their pain intensity on a trial-by-trial basis, and had their brain responses recorded using EEG. We will analyze:

• Person-to-person variations: The relationship between individual mean brain responses (N1/N2/P2 components, alpha/beta/gamma activity) and individual mean pain ratings across participants will be assessed using Bayesian multivariate linear regression. To test the repeatability of relationships, the analysis will be repeated for session 2.

• Moment-to-moment variations: The relationship between single-trial brain responses and single-trial pain ratings will be assessed, and its consistency at the group-level tested using Bayesian linear mixed effects models. To test the repeatability of relationships, the analysis will be repeated for session 2.

• Day-to-day variations: The relationship between individual mean brain responses and individual mean pain ratings averaged within the two sessions will be assessed, and its consistency at the group-level tested using a single Bayesian generalized mixed effects model.. This approach renders variability across days similar to how trial-by-trial analyses capture moment-to-moment variability.

Understanding peak alpha frequency (PAF) – Relationships to measures of lifestyle and mental well-being 

The electrical activity of the human brain is dominated by neuronal oscillations between 8 and < 13 Hz (Pernet, Garrido, et al., 2020), termed alpha rhythm. A key feature of the alpha rhythm is its peak frequency; the frequency with the highest peak in the alpha band (peak alpha frequency, PAF). PAF can be easily measured by non-invasive electro- or magnetoencephalography (EEG/MEG) recordings. In recent years, it has increasingly attracted attention as a potential clinical biomarker that is easy to apply, cost-efficient, and scalable.

The functional significance of PAF is, however, not fully clear yet. Converging lines of evidence indicate that PAF might mark fundamental neuronal processes relevant across health and disease. PAF has a strong genetic component, is heritable, and displays high intra-individual stability. It, however, changes across the lifespan, first increasing until adolescence and then slowing down throughout adulthood. Functionally, it has been related to various cognitive processes in health and disease. For instance, a faster PAF was linked to better cognitive performance and less sensitivity to pain in healthy people. In clinical populations, a slowing of PAF has been demonstrated across different diseases, including schizophrenia, dementia, and some chronic pain samples.

Together, these findings suggest that PAF might not reflect a particular cognitive process or clinical condition but more general aspects of brain function. It might, thus, be a basic neurophysiological marker related to brain health. Such a marker could serve different biomarker functions. It could indicate the risk/susceptibility to develop brain disorders. Moreover, it could be a prognostic or predictive biomarker predicting disease courses or treatment responses. Finally, it could serve as a monitoring biomarker. PAF could thus help predict, prevent, and individually treat brain disorders. Since EEG is broadly available, cost-efficient, mobile, and scalable to large numbers of people, PAF has high translational potential.

The current secondary data analysis aims to better understand the functional significance of PAF. To this end, we will reanalyze resting-state recordings of a large EEG data set in healthy human participants. Comprehensive behavioral data characterizing mental and physical well-being and lifestyle (quantifying factors like sleep quality, positive and negative emotions, body mass index, and usual alcohol consumption) will be used to predict inter-individual variations in PAF derived from eyes-closed resting-state recordings. Analyses will aim to identify relevant associations between these variables and PAF, as well as with basic demographic factors like age and gender. Probing the relationship of PAF to these basic measures will help to understand its functional role and translational potential as a clinically useful biomarker.

Resting-state EEG in depression - a systematic review and meta-analysis

Does resting state EEG activity differ between patients with depression and healthy participants? Does resting state EEG activity relate to the severity of depressive symptoms? Answering these questions will help to elucidate the potential of EEG to define biomarkers and treatment targets in depression. The study design is suitable for detecting diagnostic (i.e., differentiating patients from healthy participants) and monitoring (i.e., stratifying disease severity) biomarkers related to the BEST (Biomarkers, EndpointS, and other Tools) initiative. Moreover, both potential EEG biomarker types might be promising targets for new neuromodulatory treatment approaches such as (non-invasive) brain stimulation or neurofeedback.

Using somatosensory entrainment to modulate neuronal oscillations and the experience of pain

Sensory stimulation (e.g., visual) constitutes a simple and straightforward point of entry to the human brain, and rhythmic sensory stimulation is currently emerging as a promising non-invasive technique to induce or modulate neuronal oscillations within specific sensory brain networks and, thereby, modulate perception and behaviour. Thus, we will explore in healthy volunteers whether rhythmic somatosensory stimulation can modulate the experience of pain. We will also analyse sensory entrainment effects on oscillatory brain activity and brain connectivity in pain-relevant brain areas, including the insular, cingulate, and opercular cortex.

Using tRNS to modulate cortical excitability and the experience of pain

tRNS is an increasingly used tACS variant in which alternating currents at a mix of high frequencies (>100Hz) are applied. This approach does not aim at entraining oscillatory brain activity but at modulating cortical excitability and thereby indirectly modulating brain oscillations. Accumulating evidence shows that tRNS can induce particularly long-lasting modulations of perception and cognition. In healthy volunteers, it will be explored whether tRNS over the motor and dorsolateral prefrontal cortex can modulate the experience of pain. Moreover, the tRNS effects on oscillatory brain activity and cortical excitability assessed by TMS will be studied, and the recently established 1/f signal as a measure of cortical excitability in motor and dorsolateral prefrontal cortex, as well as in insular, cingulate, and opercular cortex.