GRAN-EEG: A ghost-residual attention network for subject-independent EEG-based objective pain recognition

Abstract

Objective pain assessment is important for supporting clinical decision-making, particularly when patients are unable to provide reliable self-reports. This study proposes GRAN-EEG, a Ghost-Residual Attention Network for subject-independent pain recognition using electroencephalography (EEG). Recordings from the PhysioPain dataset were used to classify four conditions: back pain, headache, menstrual pain and no pain. Each recording was segmented into overlapping windows and a 144-dimensional feature vector comprising statistical, temporal, spectral and Hjorth-based descriptors was extracted from each segment. GRAN-EEG integrates dense feature embedding, Ghost feature expansion, residual learning, squeeze-and-excitation recalibration and multi-head self-attention to learn discriminative EEG representations. Participant-level data separation was adopted to prevent information leakage, using a five-fold training procedure with a fixed subject-independent test set. The proposed model achieved a mean test accuracy of 84.22%, with macro-precision, macro-recall and macro-F1 scores of 85.32%, 84.21% and 84.31%, respectively. The strongest EEG-specific competitor, an EEG Transformer, achieved 80.60% mean accuracy and the performance difference remained statistically significant after Holm correction (adjusted p = 0.0426). Component-wise ablation showed that Ghost feature expansion provided the largest contribution while squeeze-and-excitation recalibration and multi-head self-attention produced additional improvements. SHapley Additive exPlanations (SHAP) identified influential descriptors distributed across multiple EEG frequency bands. Thus, GRAN-EEG provides an accurate and interpretable framework for subject-independent EEG-based pain-condition recognition.