Improving Statistical Machine Learning approaches for Time-to-Event Prediction Modelling

点击翻译按钮获取中文摘要

BackgroundThe implications of more than one chronic health condition occurring in the same individual on long-term health outcomes is often unclear. Existing studies typically focus on the association of a single condition with a single outcome and often ignore or select out individuals who might have a background of multiple long-term conditions (MLTC). Consequently, the health needs of certain groups of individuals with MLTCs may not be adequately addressed. One obstacle to studying the impact of MLTC is the number of possible combinations of health conditions. It may not be feasible to identify and recruit sufficiently large numbers of individuals with certain sets of conditions to study.To overcome these challenges, a recent approach has been to retrospectively use information held in medical record databases. Groups of individuals with similar characteristics can be identified from the database and compared with other groups to determine why certain health outcomes manifest. Computer models, based on statistical machine learning algorithms, can then be created to predict the future risk of these health outcome given individual characteristics. However, the use of this historical datasets to create prediction models needs careful handling. The data may have been acquired under a different context and/or premise to the situation in which you may be interested, and this could lead to computer models that give biased or misleading insights. In addition, complex machine learning models can lack robustness and be prone to unstable behaviour, for example, giving very different risk probabilities for two nearly identical patients.Aims & ObjectivesThis research aims to develop methodologies that will improve the robustness and validity of statistical machine learning-based prediction models that are constructed from observational data:1) To assess the robustness and stability of existing statistical machine learning approaches for time-to-event modelling,2) To develop methodology to improve aspects of the robustness and stability of statistical machine learning approaches for time-to-event modelling,3) To test the novel methodology using real-world primary care data and compare it to existing approaches.Novelty of the research methodologyStandard machine learning development focuses on the use of accuracy-related criteria to measure how well prediction models perform. However, there is increasing awareness that in real-world usage, accuracy is just one of several important criteria that determines the usefulness of a prediction model. In this research we will study the use of model training criteria that encompass considerations of the (i) four levels of model stability (as defined in Riley & Collins (2023), (ii) consistency between model versions after updating, and (iii) sensitivity to unusual data inputs.Alignment to EPSRC's strategies and research areasThis project falls within the 'EPSRC Healthcare Technologies research area' where "Optimising disease prediction, diagnosis and intervention" is one of the themes or research areas listed on this website https://www.epsrc.ac.uk/research/ourportfolio/themes/It will create new methods for analysing large real-world primary care health data sets, underpin patient-specific predictive models, and support the identification of opportunities for prevention of disease or its recurrence.CollaborationsThis project will involve a collaboration with the University of Birmingham.