Use this URL to cite or link to this record in EThOS:
Title: Vital-sign data-fusion methods to identify patient deterioration in the emergency department
Author: Santos, Mauro
ISNI:       0000 0004 7653 4885
Awarding Body: University of Oxford
Current Institution: University of Oxford
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
In the United Kingdom Emergency Departments (ED), clinical staff requires to diagnose, give treatment and discharge patients, within 4 hours of their arrival. The patients' vital signs are traditionally managed using paper Track and Trigger (T&T;) charts, prone to human error, and bedside monitors, whose alerts are often ignored. Consequently, patient deterioration might be missed at and between nurses' observation sets. This thesis has analysed data from a three stage study in the ED of the John Radcliffe Hospital, Oxford, to investigate the use of an electronic T&T; system (VitalPac) and a data-fusion system (Visensia) to help staff identify physiological deterioration in patients attending the majors area. Data was collected from a total of 10,488 ED attendances receiving standard care in stage 1, followed by two technology interventions in stages 2 and 3, respectively, for a total period of 6 months. It was shown that 9% of the observations sets, conducted on stable ED patients in stage 1, were done on unstable patients when staff was guided by VitalPac in stage 2. One of the causes might have been the increase in the Early Warning Score (EWS) completion from 52% to 100% of the observation sets. In stage 3, 35.7% of the Visensia alerts generated on continuous bedside monitor data, were deemed "technical" alerts due to data artefacts. On the other hand, clinical staff responded within 15 minutes to 85% of the "physiological" alerts. A two-stage Machine Learning (ML) architecture was proposed to fuse intermittent and continuous vital-sign data and use a sub-population novelty detection model to identify multivariate data deviating from normal physiology. This ML approach out-performed the baseline Visensia model (a population based model, applied over the continuous data), and the National EWS system (applied over the observation sets data) in detecting patients escalated to the Resuscitation area during their ED stay, on a test set of 1,070 ED attendances (AUROCs and 95% confidence intervals were 0.737 (0.623, 0.830), 0.657 (0.521, 0.755), and 0.643 (0.522, 0.749), respectively).
Supervisor: Tarassenko, Lionel ; Clifton, David Sponsor: RCUK Digital Economy Programme
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: Continuous patient vital-sign monitoring ; Novelty Detection ; Machine Learning ; In-hospital Early Warning Systems ; Gaussian Processes ; Emergency Department ; Kernel Density Estimation ; Support Vector Machines