The power of connections : mapping the behaviour of healthcare networks

As the National Health Service (NHS) in England enters its eighth decade, it faces unprecedented levels of demand for care from a patient population with increasingly complex needs. In an attempt to address these challenges, whether through centralisation of specialist services, or the introduction of Integrated Care Systems, the current organisational structure of the NHS is in transition. This thesis takes routinely collected administrative healthcare data in the form of Hospital Episode Statistics from 2011 to 2018 and applies a range of emerging network analysis and machine learning techniques to investigate the dynamics of the NHS as a complex system. Undirected patient-sharing networks based on 130 million hospital presentations from 2013-15 were found to connect 98% of all possible pairs of hospitals, but were highly clustered and were subdivided by Louvain Modularity into ten smaller patient-sharing communities. Within these patient-sharing communities, patients attended different hospitals on successive occasions over 14 million times. The Herfindahl-Hirschman index was used to identify regional differences in hospital market sizes and identified dilute healthcare markets within conurbations and areas in close proximity to multiple care providers. Random forest classifiers and alternative-specific conditional logit models were developed to predict the choices patients make in attending hospital for emergency care and are found to predict patient choice correctly on up to 87% of occasions. Markov Multiscale Community Detection, an unsupervised network clustering technique, is applied to identify data-driven catchment areas for planned orthopaedic care nationally and across primary and secondary care in conurbations. Collectively, these methods provide policymakers with a new understanding of how patients interact with hospitals, how those hospitals are connected to one another and how the broader organisational structure of the NHS may be tailored to match the complex relationships that exist between patients and providers.