EEG-Based Pain Classification via Sample Selection to Mitigate Subjective Label Bias

Quantifying pain intensity is essential for enabling personalized pain management. Recently, electroencephalography (EEG)-based approaches have been investigated to estimate pain levels, particularly for patients who are unable to communicate their pain due to cognitive or neurological impairments. However, most existing methods are trained using self-reported pain labels, which are inherently subjective. This subjectivity often leads to biased models that limit the reliability of predictions. To address this issue, we propose a novel method that incorporates reliable sample selection for EEG-based pain level classification during training. The proposed approach quantifies sample informativeness and estimates label reliability. Each sample is then assigned a priority level, and those identified as either unreliable or uninformative are excluded to enhance model robustness. We evaluate the method using EEG data from 41 participants exposed to warm, cool, and thermal grill illusion (TGI) stimuli, with pain labels collected via the Numerical Rating Scale (NRS). A 5-fold cross-validation procedure is employed to ensure robustness in both quantitative and qualitative evaluations. The proposed model achieves statistically significant improvements over baseline models in multi-class classification with 3, 6, and 10 classes. Furthermore, we demonstrate that our method generalizes well to previously unseen types of thermal stimulation, underscoring its potential for objective pain assessment in non-communicative patients. Additional analyses reveal pain-related EEG features, indicating that delta-band activity at the left and right frontotemporal electrodes (F7 and F8) is strongly associated with perceived pain intensity.