Title:  Enabling remote monitoring of joint health with signal processing and machine learning

Committee:

Dr. Omer Inan, ECE, Chair, Advisor

Dr. David Anderson, ECE

Dr. Mindy Millard-Stafford, BioSci

Dr. Alessio Meda, GTRI

Dr. Josiah Hester, IC

Abstract: Wearable sensing is a growing field that allows for frequent quantitative measures to supplement clinical visits. Conditions that affect joint health, such as injuries and arthritis, impact the daily lives of people around the world. Current diagnostic and monitoring techniques require clinical exams as well as imaging or blood tests, and as such, the evaluation of different treatments is time consuming, expensive, and can be subjective. Wearable sensing is one option to overcome these difficulties by providing a noninvasive, convenient, and less expensive alternative or supplement to traditional medicine. The objective of this work is to enable remote joint health sensing with a smart wearable knee brace using signal processing and machine learning. This work focuses specifically on the knee because it is one of the most affected and largest joints in the body. Joint acoustic emissions (JAEs), electrical bioimpedance (EBI), and kinematics are three noninvasive sensing modalities that have been investigated previously for joint health sensing and have been integrated into a wearable device. The current state of JAE analysis relies on benchtop setups and controlled recordings, and to transition to wearable sensing, accurate, generalizable, and consistent tools need to be developed. To accomplish this, a signal quality assessment algorithm for JAEs was developed and validated to overcome potential artifacts introduced when collecting data in an uncontrollable environment. A clinical study on rheumatoid arthritis (RA) was then completed using the sensing brace to quantify disease activity, showing potential for clinical adoption. The signal quality assessment algorithm was then applied to clean the JAE signals, and machine learning models were developed using JAEs, EBI, and kinematics to predict disease activity and inflammation levels. Then, to validate the use of one of the more variable measurements, JAEs, generalizability across domains was shown. Models for predicting joint health were transferred across setup, recording locations and researchers, and patient populations to demonstrate the capabilities for adoption as a biomarker. This work provides the algorithmic steps necessary to use wearable sensing for assessing joint health.