TY  - UNPB
N1  - Copyright © The Author [2022].  Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/).  Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms.  Access may initially be restricted at the author?s request.
SP  - 1
A1  - Wang, Chongyang
PB  - UCL (University College London)
M1  - Doctoral
Y1  - 2022/06/28/
TI  - Protective Behavior Detection in Chronic Pain Rehabilitation: From Data Preprocessing to Learning Model
AV  - public
N2  - Chronic pain (CP) rehabilitation extends beyond physiotherapist-directed clinical sessions and primarily functions in people's everyday lives. Unfortunately, self-directed rehabilitation is difficult because patients need to deal with both their pain and the mental barriers that pain imposes on routine functional activities. Physiotherapists adjust patients' exercise plans and advice in clinical sessions based on the amount of protective behavior (i.e., a sign of anxiety about movement) displayed by the patient. The goal of such modifications is to assist patients in overcoming their fears and maintaining physical functioning. Unfortunately, physiotherapists' support is absent during self-directed rehabilitation or also called self-management that people conduct in their daily life.
To be effective, technology for chronic-pain self-management should be able to detect protective behavior to facilitate personalized support. Thereon, this thesis addresses the key challenges of ubiquitous automatic protective behavior detection (PBD). Our investigation takes advantage of an available dataset (EmoPain) containing movement and muscle activity data of healthy people and people with CP engaged in typical everyday activities. To begin, we examine the data augmentation methods and segmentation parameters using various vanilla neural networks in order to enable activity-independent PBD within pre-segmented activity instances. Second, by incorporating temporal and bodily attention mechanisms, we improve PBD performance and support theoretical/clinical understanding of protective behavior that the attention of a person with CP shifts between body parts perceived as risky during feared movements. Third, we use human activity recognition (HAR) to improve continuous PBD in data of various activity types. The approaches proposed above are validated against the ground truth established by majority voting from expert annotators. Unfortunately, using such majority-voted ground truth causes information loss, whereas direct learning from all annotators is vulnerable to noise from disagreements. As the final study, we improve the learning from multiple annotators by leveraging the agreement information for regularization.
EP  - 200
UR  - https://discovery.ucl.ac.uk/id/eprint/10150092/
ID  - discovery10150092
ER  -