Exploring Electroencephalography for Chronic Pain Biomarkers: A Large-Scale Benchmark of Data- and Hypothesis-Driven Models

Resting-state electroencephalography (EEG) has been proposed as a scalable source of biomarkers for chronic pain, but its clinical potential remains uncertain. To systematically evaluate this potential, we benchmarked nine modeling strategies, spanning conventional machine learning with handcrafted features to state-of-the-art deep learning. Across 72 configurations of signal representations and model architectures, we trained models to predict self-reported pain intensity, using chronological age decoding as a positive control. Pain prediction performance was limited (R=0.15), with the best results achieved by conventional connectivity-based models. In contrast, age was robustly decoded from the same dataset (R=0.53), confirming technical efficacy. These findings indicate that resting-state EEG contains limited information about inter-individual differences in chronic pain intensity, making it unlikely to yield clinically actionable biomarkers in cross-sectional settings. Instead, its potential may lie in intra-individual modeling of pain dynamics, which could advance individualized mechanistic insights and more personalized treatment of chronic pain.