Explainable AI for pain perception: subject-independent EEG decoding using DeepSHAP and CNNs

<i>Objective.</i>Accurate classification of pain levels is essential for clinical monitoring, particularly in clinical populations with limited verbal communication. This study explores the feasibility of decoding pain from EEG using explainable deep learning.<i>Approach.</i>EEG signals from 50 subjects exposed to low and high pain stimuli were analyzed. A 1D convolutional neural network (CNN) was trained using leave-one-subject-out (LOSO) cross-validation. To enhance interpretability, DeepSHAP was applied to identify frequency-specific contributions of EEG features to the model's decisions.<i>Main Results.</i>The CNN achieved a classification accuracy of 95.85%, outperforming traditional classifiers (SVM, LDA, RF, etc.) on the same dataset. Explainability analysis showed that increased beta activity (14-15 Hz) was associated with high pain, while alpha (11-12 Hz) theta and delta bands correlated with lower pain states.<i>Significance.</i>This work demonstrates the potential of explainable deep learning in real-time, subject-independent pain decoding. The results support the integration of XAI techniques into EEG-based brain-computer interface (BCI) systems for objective pain monitoring.