Training clinically relevant EEG–based pain detection models: a pilot machine learning study

Assessing pain in non-communicative patients remains a critical challenge in clinical settings. This pilot study introduces a machine learning (ML) framework for pain detection using electroencephalogram (EEG) signals, leveraging methodological rigor to ensure model robustness and generalizability. EEG data were collected from a single healthy subject exposed to controlled mild pain stimuli (self-administered forearm pinches). The pipeline included signal normalization, wavelet-based feature extraction (Daubechies 4 wavelet), principal component analysis (PCA) for dimensionality reduction, and SMOTEENN to address class imbalance. A logistic regression model, optimized via Bayesian hyperparameter tuning and validated with stratified 5-fold cross-validation, achieved an accuracy of 84.62% (validation) and 75.68% (test), with precision/recall balances indicating low false-positive rates (90%/80%) and moderate sensitivity (75%/66.67%). Comparative analyses with SVM, ensemble methods, and Naive Bayes highlighted logistic regression’s superior generalization, supported by learning curves demonstrating stable convergence. Despite promising proof-of-concept results, limitations include the single-subject design and artificial pain stimulus, necessitating future validation in larger cohorts and comatose populations. This work underscores the importance of transparent ML training processes and robustness evaluations for clinical translation, advocating for methodological standards that prioritize generalizability over inflated performance metrics.