Continuous objective multimodal pain assessment sensing system (COMPASS)

• Online Access: http://hdl.handle.net/2047/D20416566

The United States Department of Health and Human Services recommended pain management as one of the five key strategies to address the current opioid crisis. Doctors need to treat patients precisely based on their pain level, which decides the usage of anesthetic on their patients before or after surgery. Getting an objective pain level of patients is crucial for both patients and doctors. Currently, most hospitals use the visual analog scale (VAS), numerical rating scale (NRS), and verbal rating scale (VRS), which are subjective and not consistent. An objective pain assessment system has been longed for by clinicians to improve the management and treatment of pain. This dissertation contains three studies, exploring different physiological modalities and their relations with pain. In the first study, twenty healthy subjects were recruited at Northeastern University. Electroencephalogram (EEG) signals were collected from all the subjects from cold pressor tests (CPTs). Paired t-tests were performed on the EEG spectral power to investigate the difference between the EEG data under non-pain and pain states. The experimental results showed that the gamma-band power increased significantly over extensive brain regions in pain condition. EEG proved to be a promising signal for pain detection. The second study utilized pupillary responses from thirty-two healthy subjects during CPTs. Eleven features were extracted from the pupillary data. This study performed a 3-class classification on extracted features, using the genetic algorithm and artificial neural network (ANN). The final classification accuracy over three classes of pain levels was 81.0% with selected features (the mean, the root mean square, and the pupillary area under the curve). Pupillary responses with machine learning algorithms had good performance on pain level classification. The third study proposed a multimodal pain assessment system, using eight physiological modalities with multi-level fusion. The data samples were categorized into four classes, No Pain, Low Pain, Moderate Pain, and High Pain. Besides EEG features extracted by a convolutional neural network, three hundred and seventy-nine features were extracted. Two ANN classifiers were utilized in this study and their results were fused at a decision level. The final classification accuracy was 87% over four pain classes. The proposed multimodal pain assessment system demonstrated accurate and efficient classification performance.--Author's abstract