Quantification of pain severity using EEG-based functional connectivity. (August 2021)

• Record Type: Journal Article
• Authors: Modares-Haghighi, P.; Boostani, R.; Nami, M.; Sanei, S.
• Is Part Of: Biomedical signal processing and control. Volume 69(2021)
• Publication Date: 2021-08
• Subjects: Brain connectivity graph -- Partial directed coherence -- Graph filtering -- Pain measurement -- EEG; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102840
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Background and objectives: The traditional pain measures are qualitative and inaccurate. Therefore, electroencephalography (EEG) signals have been recently used and analyzed to differentiate pain from no-pain state. The challenge is emerged when the accuracy of these classifiers is not enough for differentiating between different pain levels. In this paper, we demonstrate that EEG-based functional connectivity graph is remarkably changed by increasing the pain intensity and therefore, by deriving informative features from this graph at each pain level, we finely differentiate between five levels of pain. Methods: In this research, 23 subjects (mean age: 22 years, Std: 1.4) are voluntarily enrolled and their EEG signals are recorded via 29 electrodes, while they execute the cold pressure test. The signals are recorded two times from each case, while subjects press a button to annotate the EEGs into five pain levels. After denoising the EEGs, the brain connectivity graph in the Alpha band are estimated using partial directed coherence method in successive time frames. By observing the differences of connectivity graph features in different levels of pain, a bio-inspired decision tree (multilayer support vector machines) is proposed. Discriminant features are selected using sequential forward feature selection manner and the selected features are applied to the proposed decision tree. Results: Classification result for differentiating between the pain and no-pain states provides 92% accuracy (94% sensitivity and 91% specificity), while for the five classes of pain, the proposed scheme generates 86% accuracy (90% sensitivity and 82% specificity), which is slightly decreased compared to the two-class condition. Moreover, the results are evaluated in terms of robustness against noise in different signal to noise ratio levels. Comparison results with previous research imply the significant superiority of the proposed scheme. Conclusion: In this paper, we show that the elicited features from the filtered brain graph are able to significantly discriminate five different levels of pain. This is therefore the amount of co-activation between the brain regions (graph links) are significantly varied, as the pain feeling increases. Our observations are consistent with the physiological observations acquired from the images of functional magnetic resonance and magnetoencephalography.