Personalised exercise recognition towards improved self-management of musculoskeletal disorders.
Professor Nirmalie Wiratunga email@example.com
Professor Kay Cooper firstname.lastname@example.org
Doctor Stewart Massie email@example.com
Musculoskeletal Disorders (MSD) have been the primary contributor to the global disease burden, with increased years lived with disability. Such chronic conditions require self-management, typically in the form of maintaining an active lifestyle while adhering to prescribed exercises. Today, exercise monitoring in fitness applications wholly relies on user input. Effective digital intervention for self-managing MSD should be capable of monitoring, recognising and assessing performance quality of exercises in real-time. Exercise Recognition (ExRec) is the machine learning problem that investigates the automation of exercise monitoring. Multiple challenges arise when implementing high performing ExRec algorithms for a wide range of exercises performed by people from different demographics. In this thesis, we explore three personalisation challenges. Different sensor combinations can be used to capture exercises, to improve usability and deployability in restricted settings. Accordingly, a recognition algorithm should be adaptable to different sensor combinations. To address this challenge, we investigate the best feature learners for individual sensors, and effective fusion methods that minimise the need for data and very deep architectures. We implement a modular hybrid attention fusion architecture that emphasises significant features and understates noisy features from multiple sensors for each exercise. Persons perform exercises differently when not supervised; they incorporate personal rhythms and nuances. Accordingly, a recognition algorithm should be able to adapt to different persons. To address the personalised recognition challenge, we investigate how to adapt learned models to new, unseen persons. Key to achieving effective personalisation is the ability to personalise with few data instances. Accordingly, we bring together personalisation methods and advances in meta-learning to introduce personalised meta-learning methodology. The resulting personalised meta-learners are learning to adapt to new end-users with only few data instances. It is infeasible to design algorithms to recognise all expected exercises a physiotherapist would prescribe. Accordingly, the ability to integrate new exercises after deployment is another challenge in ExRec. The challenge of adapting to unseen exercises is known as open-ended recognition. We extend the personalised meta-learning methodology to the open-ended domain, such that an end-user can introduce a new exercise to the model with only a few data instances. Finally, we address the lack of publicly available data and collaborate with health science researchers to curate a heterogeneous multi-modal physiotherapy exercise dataset, MEx. We conduct comprehensive evaluations of the proposed methods using MEx to demonstrate that our methods successfully address the three ExRec challenges. We also show that our contributions are not restricted to the domain of ExRec, but are applicable in a wide range of activity recognition tasks by extending the evaluation to other human activity recognition domains.
WIJEKOON, A. 2021. Personalised exercise recognition towards improved self-management of musculoskeletal disorders. Robert Gordon University, PhD thesis. Hosted on OpenAIR [online]. Available from: https://doi.org/10.48526/rgu-wt-1358224
|Deposit Date||Jun 8, 2021|
|Publicly Available Date||Jun 8, 2021|
|Keywords||Exercise recognition; Human activity recognition (HAR); Personalised activity recognition; Exercise therapy; Self-management of care|
WIJEKOON 2021 Personalised exercise recognition
Copyright: the author and Robert Gordon University
You might also like
Personalised meta-learning for human activity recognition with few-data.
Clood CBR: towards microservices oriented case-based reasoning.
Learning to compare with few data for personalised human activity recognition.
Heterogeneous multi-modal sensor fusion with hybrid attention for exercise recognition.