Pain Assessment with Electrodermal activity signals by Machine Learning Applications

Objective and automated pain detection has been one of the key concerns of clinical researches for many years. The modern techniques for automated pain detection are based on machine learning methods that distinguish pain condition from no-pain condition mainly relying on the information extracted from heart rate, electromyography, blood volume pressure and recorded video. Moreover, recent studies based on the statistical analyses of the electrodermal activity (EDA) showed that EDA carry salient information about different pain phases in patients. This is mainly due to the fact that EDA signals contain rich information about autonomic nerves system that can identify different stress levels in patients induced due to varying pain levels. Due to noninvasiveness and portability of EDA measurement systems, a pain detection method based on EDA is highly valuable for clinical settings to quickly, accurately and objectively identify the pain levels of patients. In this thesis, we present a novel and highly accurate pain level detection algorithm based on EDA. Specifically, with the aim of distinguishing pain from pain-free conditions with EDA signals, we develop a comprehensive machine learning framework. In this framework, we employ timescale decomposition (TSD) to extract salient statistical features from EDA signal associated with pain. TSD uniquely disintegrates a signal in time domain to capture short and long-term changes over the signal. We characterize these changes in the EDA signals through the computation of mean, variance and entropy over all the decomposed segments of the EDA. We use these statistical and information theoretic identities to form feature vectors from the EDA. The dimensions of the feature vectors are then reduced through principle component analysis. In order to develop a pain detection method, these reduced features are then used in three different classifiers, linear discriminant analysis (LDA) and support vector machines (SVMs) with linear and radial basis function (RBF) based kernels. We test our algorithm using a dataset which is obtained through our collaboration with University of California San Diego Medical School. The EDA data in this dataset are measured from children who had undergone a laparoscopic appendectomy. Particularly, we aim to identify two phases of pain including acute and ongoing pain. Ongoing pain is as the name suggests in pain which is continuous in a period of time after the surgery while acute pain is defined as the pain induced as a result of a press in the surgical area by the research staff. Our results show that, using linear SVM, acute pain is differentiable with an accuracy of 83.33\% (sensitivity=86.67\%, specificity=80\%) while ongoing pain is distinguishable with an accuracy of 87.5\% (sensitivity=87.5\%, specificity=87.5\%) using LDA. Requiring less setup complexity than the existing methods, these results are comparable to the state-of-the-art pain detection method