Machine learning digital biomarkers from unconstrained real-world movement data in neurological disorders
Author(s)
Auepanwiriyakul, Chaiyawan
Type
Thesis
Abstract
Clinical observation and monitoring of patients with neurological conditions is one of the most essential tasks concerning managing post-neurological condition recovery. Medical professionals need methods to quantify how often patients are ambulant and use their limbs in daily activities to monitor recovery and tailor treatments. Unfortunately, current standards of care for clinical observation and behavioural monitoring rely on intermittent and subjective manual observations, artificial motor tasks, and self-reports, which both suffer from bias and incomplete data. Wearable technologies and machine learning algorithms offer a more objective alternative for continuous monitoring of behaviour. However, most machine-learning models are trained on wearable motion data collected from healthy/chronic disease populations during artificial motor tasks, which exhibit vastly different motor behaviour compared to free-living motor tasks.
In this thesis, we develop, deploy, and validate a wearable system and machine learning algorithm that can continuously monitor real-world behaviour directly in a population of patients in an acute stroke ward. Using a simple movement exercise experiment, we validate our wearable system against two existing gold-standard systems and find that our system performs well, achieving a maximum RMSE of 2.29±0.09 and a minimum R2 of 0.49±0.02, compared to the current gold-standard approaches. Additionally, we validate our system backend and show that our data is highly structured and would benefit significantly from an SQL-type database...
In this thesis, we develop, deploy, and validate a wearable system and machine learning algorithm that can continuously monitor real-world behaviour directly in a population of patients in an acute stroke ward. Using a simple movement exercise experiment, we validate our wearable system against two existing gold-standard systems and find that our system performs well, achieving a maximum RMSE of 2.29±0.09 and a minimum R2 of 0.49±0.02, compared to the current gold-standard approaches. Additionally, we validate our system backend and show that our data is highly structured and would benefit significantly from an SQL-type database...
Version
Open Access
Date Issued
2023-03-31
Date Awarded
2024-02-01
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Faisal, Aldo
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)