Wearable sensors and machine learning for tracking movement, deterioration, and recovery in acute in-patients
File(s)
Author(s)
Waibel, Sigourney
Type
Thesis
Abstract
Background— In hospital and clinical trials, healthcare professionals need to assess in-patient movement to monitor improvement and deterioration. Current means by which this is achieved involve intermittent, manual clinical assessments that are inefficient. By contrast, wearable motion sensors analysed with ML may provide an automated and continuous method to track patients’ natural movement and relate this to clinical status. However, to apply these technologies, we need to develop a robust data acquisition system and validate this in target clinical populations. This thesis describes the methodology and proof-of-concept data of a novel wearable motion sensor system applied to natural movement monitoring of acute in-patients. The aims are to establish the accuracy and end-user acceptability of wearable motion tracking technology, the validity of such a system to recognise natural activity, and its potential utility for estimating disability and identifying the risk of deterioration.
Methods—We built a patient monitoring system which combines wearable motion sensors and machine learning models. We compared smartwatch sensors’ raw motion signal (Acceleration and Angular Velocity) against optical motion tracking and research-grade sensors in a lab-based movement task in 12 healthy controls. We evaluated the models’ accuracy for recognising natural activity against human ground-truth activity labels in 123 acute in-patients. We validated the models’ performance for predicting in-hospital deterioration in 195 acute in-patients against ground truth clinician ratings. We validated the models’ performance for predicting functional dependence (modified Rankin Score >2) in 195 acute in-patients against ground truth clinician ratings. We surveyed the perceptions and attitudes regarding wearables and artificial intelligence in 166 in-patients and 30 healthcare professionals following a trial in which in-patients used the system from 9:00-18:00...
Methods—We built a patient monitoring system which combines wearable motion sensors and machine learning models. We compared smartwatch sensors’ raw motion signal (Acceleration and Angular Velocity) against optical motion tracking and research-grade sensors in a lab-based movement task in 12 healthy controls. We evaluated the models’ accuracy for recognising natural activity against human ground-truth activity labels in 123 acute in-patients. We validated the models’ performance for predicting in-hospital deterioration in 195 acute in-patients against ground truth clinician ratings. We validated the models’ performance for predicting functional dependence (modified Rankin Score >2) in 195 acute in-patients against ground truth clinician ratings. We surveyed the perceptions and attitudes regarding wearables and artificial intelligence in 166 in-patients and 30 healthcare professionals following a trial in which in-patients used the system from 9:00-18:00...
Version
Open Access
Date Issued
2022-07-20
Date Awarded
01/11/2023
License URL
Advisor
Bentley, Paul
Faisal, Aldo
Sponsor
National Institute for Health Research (Great Britain)
Publisher Department
Brain Sciences
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)