Classification of chronic pain and spinal cord stimulation response using machine learning in magnetoencephalography
Bart Witjes1, Martijn P A Starmans2,3, Frank J P M Huygen1
1Center for pain Medicine, Department of Anesthesiology, Erasmus University Medical Center, Rotterdam, the Netherlands.
Plos One
|December 5, 2025
Summary
Magnetoencephalography (MEG) data shows potential as a biomarker for chronic pain, achieving 76% classification accuracy. However, its correlation with pain intensity in spinal cord stimulation patients requires further research.
Area of Science:
- Neuroscience
- Biomarkers
- Machine Learning
Background:
- Chronic pain is complex, encompassing physical, psychological, emotional, and social factors, making objective quantification challenging.
- Magnetoencephalography (MEG) is increasingly explored for identifying pain biomarkers due to its ability to capture brain activity.
- This study investigates MEG data's potential as a biomarker for chronic pain and to assess spinal cord stimulation (SCS) treatment efficacy.
Purpose of the Study:
- To explore the utility of magnetoencephalography (MEG) data in identifying chronic pain biomarkers.
- To apply machine learning models to MEG data for classifying chronic pain.
- To quantify the effect of spinal cord stimulation (SCS) treatment using MEG-derived biomarkers.
Main Methods:
- Recruited 25 chronic pain patients, 25 SCS patients, and 25 pain-free controls.
- Recorded resting-state MEG data and extracted spectral features (theta, alpha, beta, low-gamma power; alpha peak frequency; alpha power ratio).
- Utilized automated machine learning for classification and regression models based on spectral features.
Main Results:
- Achieved 76% accuracy in classifying chronic pain patients versus controls using theta power and alpha power ratio.
- The classification model output showed a weak correlation (Spearman's rho = 0.12) with self-reported pain scores in SCS patients.
- Regression models based on all participants' pain scores yielded correlations between 0.27 and 0.41.
Conclusions:
- MEG data, specifically theta power or alpha power ratio, shows promise for classifying chronic pain with 76% accuracy.
- The current model's poor correlation with SCS patient pain scores suggests limitations in capturing treatment effects.
- Future research should incorporate a wider range of input features and outcome parameters for improved pain assessment and treatment monitoring.
