An EEG based hierarchical classification strategy to differentiate five intensities of pain

• Published in: Expert systems with applications Vol. 180; p. 115010
• Main Authors: Afrasiabi, Somayeh, Boostani, Reza, Masnadi-Shirazi, Mohammad Ali, Nezam, Tahereh
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 15.10.2021; Elsevier BV
• Subjects: Alpha waves; Bayesian Optimization; Data distribution; Decision trees; EEG; Electroencephalography; Feature extraction; Pain; Pain measurement; Physiological based classifier; Support vector machines; Bayesian Optimization; Data distribution; Alpha waves; Physiological based classifier; EEG; Pain measurement
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.115010

•Objective Pain Measurement is necessary for patients who cannot express their pain.•The EEGs in alpha band for no-pain state have normal distribution.•As the pain severity increases, the PDF deviates from Gaussian distribution.•Five pain states are recognized by specifying the changes in the PDF of alpha band.•Pain features are classified using a proposed bio-inspired hierarchical classifier. Past research emphasized on revealing the pain in different bands of electroencephalogram (EEG) including alpha band. In this study, we proposed an accurate and robust manner to differentiate pain intensities by deeply characterizing the alpha band in terms of distribution, spectrum and complexity changes in response to five different intensities of pain. Here, 44 subjects executed the Cold Pressor Task (CPT) and experienced five defined levels of pain while their EEGs were recorded via 34 silver channels. After de-noising and filtering the EEGs through the alpha band, 12 informative features were extracted from each channel in successive time frames. Since none of the features could discriminate the five classes, we applied the Kruskal-Wallis test to the features for observing their distribution in differentiating two or more classes. According to this result, we designed a decision tree classifier, where a Bayes optimized support vector machine (BSVM) was selected in each decision node. Sequential forward selection was applied in order to customize a subset of features for each BSVM. Our results provided 93.33% accuracy over the five classes and also generate 99.8% accuracy for separating pain and no-pain classes, which is statistically superior (P < 0.05) to state-of-the-art methods over our collected dataset.