A Novel Classification Strategy to Distinguish Five Levels of Pain Using the EEG Signal Features

Abstract

Research studies have tried to extract pain-related features from electroencephalogram(EEG) signals for quantitative measuring of pain. In this study, we go one step further to measure three/five levels of pain by proposing efficient EEG processing steps in conjunction with a new classification strategy. 24 healthy subjects voluntarily performed the cold pressor test while their EEGs were recorded. First, the EEGs were decomposed by independent component analysis and the artifact sources were removed. Among the remained sources, pain-related sources, were chosen according to an adopted information criterion. Next, the EEGs were reconstructed by projecting back the selected sources. Then, grand average brain maps of train subjects were estimated for each pain level over the Alpha(8-12 Hz) and Delta(0.5-4 Hz) bands. By tracing the brain maps' changes over different pain levels, the structure of the proposed decision tree was formed. To enrich the feature set, we also extracted other EEG features. For each decision node, a specific subset of features was selected by sequential forward selection method. Considering k-nearest neighbor(KNN) as the decision marker,the classification accuracies for the three and five pain levels was determined 80 ± 5 and 60 ± 5 percent, respectively while by choosing support vector machine(SVM), the results improved up to 83 ± 5 and 62 ± 6 percent,respectively.