A Physiological-Inspired Classification Strategy to Classify Five Levels of Pain

Current research on quantitative pain measurement using the electroencephalogram (EEG) signals showed a promising result just on classifying pain from no-pain states. In this paper, we go one step further introducing pain-level dependent EEG features as well as proposing a physiologically-inspired hierarchical classifier to provide promising results for differentiating five classes of pain. In this research, forty four subjects were voluntarily enrolled, each of whom executed the Cold-Pressor Test (CPT), while their EEGs were simultaneously recorded. We filtered the EEGs through the alpha band and elicited meaningful features to reveal the behavior of signals in terms of distribution, spectrum and complexity at each pain state. To assess the susceptibility of the features in classifying one/group of classes against others, Kruscall-Walis test was applied to give a clue in order to construct the structure of our decision tree, where a Bayesian optimized support vector machine (BSVM) was trained at each node. After arranging the tree, sequential forward selection (SFS) was applied to select a customized subset of features for each node. Our results provide 93.33% accuracy for the five classes and also generate 99.8% for pain and non-pain classes, which is statistically superior (P<; 0.05) to state-of-the-art methods over the same dataset.