Introduction: Chronic pain is a widespread and complex condition. Spinal Cord Stimulation (SCS) offers effective pain relief in a portion of patients suffering from chronic pain, although its underlying mechanisms of action remain unclear and may differ between tonic and burst stimulation paradigms. Brain connectivity analysis can help reveal how chronic pain and SCS affect communication between brain regions. Magnetoencephalography (MEG) is particularly suited for this due to its high temporal resolution. Graph theory enables modelling of whole-brain networks, and Graph Neural Networks (GNNs), a deep learning approach designed for graph-structured data, is well-suited for distinguishing specific connectivity patterns within complex network structures. While promising, GNNs have not yet been applied to SCS or chronic pain. Furthermore, beyond classification, explainability approaches allow insights into which graph substructures drive GNN model's decisions.

Aim: The overarching aim of my exploratory study was to develop and train a GNN model based on MEG data from patients with chronic pain with SCS, to identify differences in brain networks during stimulation ON and OFF.

Methods: Resting-state MEG data were collected from 22 chronic pain patients receiving SCS, recorded in two institutes. A cyclic stimulation protocol (1 min ON, 1 min OFF) was used.
Brain connectivity graphs were constructed using the phase lag index as functional connectivity metric, and features for each brain region were derived from the power spectral density. Graph datasets were created per frequency band, stimulation paradigm (tonic and burst), and recording institute. Separate GNN models were trained to classify stimulation ON and OFF states, and explainability techniques were implemented to unravel the key graph substructures driving the model's classification decisions.

Results: GNN models accurately classified stimulation states, especially using full-band, beta, and gamma graphs (accuracies: 0.99, 0.97, 0.99). Delta, theta, and alpha bands showed lower performance (accuracies: 0.76, 0.80, 0.77). Model performance was consistent across tonic and burst SCS paradigms and both recording sites (accuracies: 0.97, 0.98, 0.99, 0.97), however, performance across paradigms showed inconsistencies. Specifically, the model trained on tonic SCS and tested on burst SCS recordings showed a cross-paradigm accuracy of only 0.69. The GNN model achieved cross-site accuracies of 0.81 and 0.87 across datasets from the recording institutes, demonstrating consistent performance across patient cohorts. Furthermore, the explainability analysis outcomes highlighted several pain-related brain regions as key substructures in the graph for distinguishing stimulation ON and OFF states.

Discussion: This study introduces GNNs as a novel method for decoding brain network dynamics in chronic pain patients with SCS. The classification results and node-level explainability suggest that pain-processing regions are modulated by SCS. The cross-paradigm accuracy suggests that burst SCS only partially captures the features of tonic SCS, possibly indicating that burst SCS engages a more widespread brain network. However, interpretation of the findings is limited by the small sample size, inter-patient variability, and the inability to separate chronic pain effects from stimulation effects. Nevertheless, this framework offers a promising direction for application of GNNs for unravelling complex brain network dynamics in chronic pain and SCS. Future studies should focus on expanding this framework by utilizing GNN models to classify SCS treatment effectiveness, potentially providing more insights into the brain regions and connectivity patterns that are most predictive of treatment success.