Use this URL to cite or link to this record in EThOS:
Title: Modelling pathways to diagnosis of breast conditions
Author: Saville, Christina
ISNI:       0000 0004 7234 2915
Awarding Body: University of Southampton
Current Institution: University of Southampton
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
This thesis describes how logistic regression and discrete-event simulation (DES) can be combined to predict patient risk and evaluate the potential operational impact of implementing risk-based pathways. We test whether using diagnostic information provided by non-specialists to plan diagnostic tests offers benefits in operational performance. We demonstrate our approach on an application area in breast diagnostics with data from the Whittington Hospital breast diagnostic clinic. Specifically we assess whether GP referral information is complete and accurate enough for use in predicting the risk of an abnormal result (i.e. an abnormality being detected from mammogram, ultrasound or biopsy). The construction of a unique dataset for this purpose is described; it links GP referral information to in-clinic tests and results. This dataset is used to develop two alternative logistic regression scorecards that predict a patient’s risk of abnormal breast diagnostic results from their GP referral data (n=179). The simple scorecard uses two referral characteristics while the full scorecard uses seven. It is usual to base the decision of where to set the cut-off score between low and high risk patients on a scorecard’s predictive performance. In contrast, we show how a discreteevent simulation can be used to optimise the cut-off in terms of operational performance. In our example, the performance measure is the daily average proportion of patients’ time at the clinic that adds value, called the clinic efficiency. We simulate the potential impacts of introducing the following risk-based pathways. High-risk patients are sent straight for imaging tests and then to a clinician for their results. Low-risk patients are sent to a clinician first (as today) who decides whether imaging is needed. The set of labels that determines a patient’s progress through the simulation is modelled empirically for the simple scorecard, since all possible label combinations are present in our sample. However for the full scorecard this is not the case, so using the empirical distribution is not appropriate. Instead we introduce a novel method using Poisson loglinear models to generate representative sets of patient labels.
Supervisor: Smith, Honora ; Bijak, Katarzyna Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available