691: Identifying Pain-Related Cortical Adaptations in Healthy Subjects Using Machine Learning

Topic: Experimental (Non-clinical) Human Studies
Subtopic: Non -specific to any diagnosis, Electrophysiology/Clinical neurophysiology (e.g. microneurography, EEG, MEG, OPMs, TMS), Artificial Intelligence/Machine Learning, Pain models (e.g. saline injection, capsaicin, UVB, HFS, incision, exercise, itch)

Background and Aims: Reduced adaptive responses of the nervous system, i.e., neuro-adaptability, in specific brain regions have been linked to the development of chronic pain, suggesting that restricted cortical flexibility may increase vulnerability to persistent pain. Transcranial Magnetic Stimulation combined with Electroencephalography (TMS-EEG) enables direct assessment of cortical excitability and connectivity. However, consensus is lacking on which TMS-EEG measures best capture neuroadaptive changes associated with pain. This study applies machine learning (ML) to extract complex patterns within the TMS-EEG data beyond classical data analysis approaches. The goal is to establish reliable markers of cortical neuro-adaptability during and following experimental tonic pain.

Methods: Healthy subjects (aiming for N=120) will undergo TMS-EEG at baseline, 30 minutes after induction of tonic experimental pain through topical capsaicin, and again 30 minutes after capsaicin removal. TMS-EEG will be recorded during primary motor cortex (M1) stimulations guided using neuronavigation. The recordings will be preprocessed to extract cortical excitability and connectivity measures across pain-relevant regions. These features will be used in ML classifiers including Support Vector machines (SVM) and Random Forrest to distinguish different binary pain states. Feature-importance algorithms such as SHapley Additive exPlanations (SHAP) will be applied to identify the most informative biomarkers detecting pain-related changes.

Results: Initial data analysis performed on the training and validation splits of the currently available data showed that the SVM model achieved a classification accuracy of 87.5% for baseline vs. pain. The SHAP algorithm identified the Inter-Trial Coherence (ITC) for the parieto-occipital region as the primary feature driving the classification. The local mean field power (LMFP) at the central-parietal left was the second most informative feature. The initial data analysis demonstrates feasibility of the study, which can be expected to be improved with a full data set to be presented at the IASP World Congress 2026.

Conclusions: The preliminary results obtained using the currently available dataset demonstrate the study’s feasibility and suggest that excitability and connectivity in parietal regions following TMS are important markers of neuro-adaptability to experimental tonic pain.

References: 1. Graven-Nielsen T. Mechanisms and manifestations in musculoskeletal pain: from experimental to clinical pain settings. Pain. 2022 Nov;163(S1):S29–45
2. Tremblay S, Rogasch NC, Premoli I, Blumberger DM, Casarotto S, Chen R, et al. Clinical utility and prospective of TMS–EEG. Clin Neurophysiol. 2019;130(5):802–44.
3. De Martino E, Casali A, Casarotto S, Hassan G, Couto BA, Rosanova M, et al. Evoked oscillatory cortical activity during acute pain: Probing brain in pain by transcranial magnetic stimulation combined with electroencephalogram. Hum Brain Mapp. 2024;45(6):e26679.
4. Lötsch J, Ultsch A. Machine learning in pain research. Pain. 2018;159(4):623–30.

Presenting Author: Jacob Sørensen
Submitting Author: Jacob Sørensen

Senior Author: Thomas Graven-Nielsen

IASP 2026 Global Year theme of Neuropathic Pain: No

• Bruno Andry Nascimento Couto: MSc - Aalborg University
• Laura Sirucek: PhD - Aalborg University
• Nahian Chowdhury: PhD - University of New South Wales
• Carolina Queiroz: MSc - Aalborg University
• Daniel Ciampi de Andrade: Prof., MD - Aalborg University
• Thomas Graven-Nielsen: Prof, MD - Aalborg University