Signature patterns in digital representations of health before acute illness
File(s)
Author(s)
Post, Benjamin
Type
Thesis or dissertation
Abstract
Emergency hospital admissions are major health events linked to high mortality rates, resource consumption and costs. The early identification of at-risk patients could enable medical intervention to prevent this outcome. Previous studies have used electronic health records to create emergency hospital admission predictors, but have exclusively used tabular data which captures only a static understanding of a patient. This thesis explores how dynamic temporal patterns of health activity can be incorporated into models to predict emergency hospital admission and death.
The Secure Anonymised Information Linkage Databank stores electronic health record data for the majority of the population of Wales. These data were transformed into simplified time series representations based exclusively on their timestamp data. Three different approaches were tested to capture time series patterns which were incorporated into emergency hospital admission prediction models. These techniques were tailored to provide clinically useful predictions, which could be adapted to a variety of datasets. Model-specific explainability approaches were developed to improve clinical utility and trust in these predictions. An ethical experiment, based on a utilitarian framework, was conducted to explore the moral
justification for this task. Human-understandable patterns of primary care activity were extracted from these health data and shown to be associated with risk of emergency hospital admission and death. These temporal patterns were incorporated into prediction models that matched the performance of established techniques, whilst requiring significantly simpler data processing. A selection of bespoke explainability techniques provided clinical context to these predictions. The research presented in this thesis demonstrates that highly simplified representations of electronic health record data possess temporal patterns that are predictive of emergency hospital admission and death. These novel approaches require simple timestamp data, making them suitable for deployment on heterogeneous datasets. A focus on clinical utility and explainability make these models ideal for real world deployment.
The Secure Anonymised Information Linkage Databank stores electronic health record data for the majority of the population of Wales. These data were transformed into simplified time series representations based exclusively on their timestamp data. Three different approaches were tested to capture time series patterns which were incorporated into emergency hospital admission prediction models. These techniques were tailored to provide clinically useful predictions, which could be adapted to a variety of datasets. Model-specific explainability approaches were developed to improve clinical utility and trust in these predictions. An ethical experiment, based on a utilitarian framework, was conducted to explore the moral
justification for this task. Human-understandable patterns of primary care activity were extracted from these health data and shown to be associated with risk of emergency hospital admission and death. These temporal patterns were incorporated into prediction models that matched the performance of established techniques, whilst requiring significantly simpler data processing. A selection of bespoke explainability techniques provided clinical context to these predictions. The research presented in this thesis demonstrates that highly simplified representations of electronic health record data possess temporal patterns that are predictive of emergency hospital admission and death. These novel approaches require simple timestamp data, making them suitable for deployment on heterogeneous datasets. A focus on clinical utility and explainability make these models ideal for real world deployment.
Version
Open Access
Date Issued
2024-05-17
Date Awarded
2024-10-01
Copyright Statement
Creative Commons Attribution NonCommercial NoDerivatives Licence
Advisor
Faisal, Aldo
Brett, Stephen
Sponsor
UK Research and Innovation
Grant Number
EP/S023283/1
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)
