Development of Bayesian network methodologies for applications at the socio-medical interface

Abstract

The integration of biomedical and social science data enables systems-based analyses through advanced statistical models. Among these, Bayesian networks (BNs) have emerged as a powerful tool for modelling complex interrelations among variables. This dissertation explores several methodological dimensions of BNs and their applications to real-world socio-medical data. First, missing data in BN structure learning is tackled by comparing multiple imputation by chained equations (MICE) with structural expectation-maximization (SEM). Simulation studies across various missingness mechanisms reveal that SEM outperforms MICE, providing more accurate structures. A real-world application to the United States Health and Retirement Study (HRS) underscores the effectiveness of SEM in uncovering associations between socio-demographic factors and chronic conditions. Second, the performance of four BN scoring functions – AIC, BIC, BDe, and log-likelihood is assessed, particularly with polychotomous data. Extensive simulations offer practical guidance on their application in various scenarios, tailored to different study objectives. A method for optimizing the BDe scoring function is developed to enhance robustness of BNs. Third, causal BNs are applied to model complex drivers of multi-drug resistance (MDR) in urinary tract infection patients in East Africa, underscoring the necessity of systems-based approaches. The BN reveals that key demographic and environmental factors significantly influence MDR. These findings provide actionable insights for policymakers. Lastly, to improve interpretability of BNs, a similarity score is developed to quantify strength and direction of interactions among categorical variables. It shows superior performance compared to traditional influence scores, particularly with higher level categorical data. A case study using HRS data demonstrates its practical application, highlighting the method’s potential in integrated social science and biomedical research. Overall, this dissertation advances the application of BNs in interdisciplinary research, offering methodological innovations for handling missing data, optimizing scoring functions, addressing complex health issues like MDR, and enhancing interpretability of interaction patterns in complex socio-medical data.

Date of Award	3 Dec 2024
Original language	English
Awarding Institution	University of St Andrews
Supervisor	V Anne Smith (Supervisor) & Katherine Lisa Keenan (Supervisor)