EEG-based Pain Level Classification Using Time-Frequency Features and Deep Learning

Nociception refers to the sensory mechanism activated by external stimuli that elicit physiological responses interpreted as pain. The perception of pain is inherently subjective, influenced by individual emotions and behaviors. Although analog pain scales have been created to address this variability, researchers are now pursuing more objective assessment methods. Recent technological advancements are exploring biomarkers to render pain assessment more quantitative and less reliant on subjective interpretation by using biosignals as the primary data source. This study presents a method for discriminating between low and high-pain classes using a public database of biosignals collected from multiple subjects exposed to a transcutaneous laser. The method leverages time-frequency information by applying a wavelet transform to EEG signals and training an inter-subject classifier. This classifier, based on a Transformer architecture, is evaluated using a leave-one-out cross-validation scheme. The experiments demonstrate that it is possible to distinguish between high and low pain states. Given the dataset information, two experiments were conducted considering two ways of labeling the data: based on laser intensity or reaction time. The best results were obtained with the second approach, where 16 out of 51 subjects achieved classification accuracies above 70%. In contrast, with the first scheme, only 4 subjects reached a similar percentage.