External validation of EEG-based machine learning models for continuous pain prediction

Machine learning (ML) has been used to predict subjective pain intensity from electroencephalographic (EEG) data. However, few ML models for pain assessment have been externally validated, particularly for continuous pain prediction. We externally validated ML regression models using multi-stage validation to predict pain intensity from single-trial oscillatory EEG features. Ninety-one participants were recruited across three samples. Sample one (n = 40) was used for model development, sample two (n = 51) for cross-subject external validation, and sample three (n = 25), comprising additional recording sessions from sample one participants, for within-subject temporal external validation. Pneumatic pressure stimuli were delivered at 10 graded intensity levels. Single-trial EEG time-frequency features were used to train a Random Forest (RF) model and a long short-term memory (LSTM) network to predict pain intensity. Both models outperformed a random prediction model, with the RF achieving mean absolute errors (MAE) of 19.59, 21.29, and 18.90 for internal validation, cross-subject external validation, and within-subject external validation, respectively. However, neither model outperformed a baseline dummy model predicting the mean training-set rating. We also trained a RF classifier to distinguish low and high pain trials, achieving internal and external validation accuracies up to 64% and 58%, respectively. Overall, our results indicate that predicting continuous pain ratings from oscillatory EEG changes using ML remains highly challenging. Classification models demonstrated above-chance performance across validation samples but remained below clinically meaningful thresholds. Further methodological improvements, particularly the integration of composite measures, are required before ML can achieve clinical utility in pain assessment. PERSPECTIVE: This study provides a rigorous evaluation of EEG-based machine learning for pain assessment, demonstrating that while continuous pain prediction remains unreliable, discrete pain states can be reasonably decoded. These findings represent an important step toward clinical validation and establish realistic performance benchmarks for developing clinically applicable, brain-based pain assessment tools.