With ageing populations and the increasingly effective management of many diseases, more people worldwide are living with Multiple Long-Term Conditions (MLTC). People with MLTC represent a highly heterogeneous population, creating challenges to the design and implementation of interventions. As a result, identifying distinct patterns, in the form of clusters of similar diseases or individuals with similar patterns of diseases, may enable personalised risk prediction and stratification, and more efficient health service design. Using general practice (GP) electronic health record (EHR) data from the Clinical Practice Research Datalink, this thesis integrates clinical epidemiological approaches with computational methods, including algorithms derived from natural language processing capable of analysing disease sequences.

Focusing on a set of 212 long-term conditions, the thesis first addresses the challenges
associated with using EHRs to delineate acute from chronic conditions to define MLTC, and the impact of factors external to patients, such as GP practice and financial incentives on the frequency of diagnostic codes. I develop and evaluate a pipeline for generating clusters of diseases and find clinically interpretable clusters corresponding to both well-established and novel patterns. Although clusters correspond to meaningful groups, they were highly heterogeneous in associations of each disease within them with future healthcare utilisation and mortality. Clusters also performed poorly at predicting patient outcomes compared with using information on a person’s individual diseases. Finally, I develop and compare methods of generating vector representations of people based on their chronological order of disease development. Although these perform well for predicting future outcomes, the clusters of people generated from them were complex to interpret and offered limited insights from a health service perspective.

Collectively, these findings provide detailed insights into methods of generating clusters and the applications and limitations of using clusters to understand patient healthcare journeys and outcomes in the context of MLTC.