Publication Date

2020

Document Type

Dissertation

Committee Members

Tanvi Banerjee, Ph.D. (Advisor); Krishnaprasad Thirunarayan, Ph.D. (Committee Member); Michael L. Raymer, Ph.D. (Committee Member); Nirmish Ramesh Shah, Ph.D. (Committee Member)

Degree Name

Doctor of Philosophy (PhD)

Abstract

Sickle cell disease (SCD) is an inherited red blood cell disorder that can cause a multitude of complications throughout a patient's life. Pain is the most common complication and a significant cause of morbidity. Since pain is a highly subjective experience, both medical providers and patients express difficulty in determining ideal treatment and management strategies for pain. Therefore, the development of objective pain assessment and pain forecasting methods is critical to pain management in SCD. On the other hand, the rapidly increasing use of mobile health (mHealth) technology and wearable devices gives the ability to build a remote health intervention system for SCD. Hence, the objective of this study is to leverage machine learning techniques, mHealth, and wearable devices together to improve pain management in SCD in both clinical and remote environments. First, we developed an objective pain assessment model based on clean physiological measurements collected from Electronic Health Records (EHRs). Specifically, we used six objective physiological measures in EHRs as features to estimate pain scores based on an 11-point pain rating scale and other pain rating scales. Overall, our preliminary machine learning models show that subjective pain scores can be predicted with objective physiological signals with promising results. Second, we designed a regression-based pain assessment model using noisy physiological and body movement data obtained from wearable devices, patient-reported pain scores from our self-developed mobile app, and nursing-obtained pain scores. The performance of the proposed model is comparable to the model learned with EHRs. We also compared the performance of the regression model and the classification model on the pain intensity estimation problem. Third, we further implemented an ensemble feature selection method to select the most robust and stable features in pain estimation to better understand pain. With robust feature selection and stacked generalization of different regression models, we were able to obtain a more compact and generalizable pain assessment model. Finally, we applied the self-supervised learning method to build a pain forecasting system with limited pain value labels. Our system outperformed the model trained in a purely supervised manner. Such a pain forecasting system would permit timely and adequate pain relief medication usage and other pain treatment plans.

Page Count

113

Year Degree Awarded

2020

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License.

ORCID ID

0000-0002-7550-0662


Share

COinS