Machine learning for the detection of clinical deterioration on hospital wards

Assessment of physiological instability preceding adverse events on hospital wards has been previously investigated through clinical early warning score (EWS) systems. Such systems often rely on routinely collected vital-sign data, such as the National Early Warning Score system that is currently deployed across the United Kingdom. In this thesis, we propose improved and generalisable risk assessment algorithms, that can continuously alert for deterioration preceding the composite outcome of unplanned intensive care unit admission, cardiac arrest, and mortality. We develop and validate three statistical and machine learning models using large-scale datasets from two independent centres: Oxford University Hospitals and Portsmouth NHS Hospitals within the HAVEN database. The retrieved datasets include patient demographics, vital signs, laboratory tests, and data of the occurrence of any adverse events. The first model is the Age- and Sex- specific Early Warning Score (ASEWS), which was derived from statistical distributions of vital signs. The work suggests that accounting for age-related vital-sign changes can more accurately detect deterioration in younger patients (16-45 years old). We also propose the Deep Early Warning Score (DEWS), which consists of an end-to-end attention-based deep learning architecture that processes vital-sign time-series data. The vital-sign data is initially modelled using Gaussian Process Regression. The results suggest that deep learning improves the detection of clinical outcomes by recognising complex patterns in the data. The final model is the information Fusion in a multi-modal Early Warning System (iFEWS), which incorporates additional information about the patient, such as results of laboratory tests or the first diagnosis assigned at admission. In this framework, representation learning and continued learning within a multi-modal system improved the performance of detecting deterioration. All of our proposed models achieved a better performance than the state-of-the-art clinical EWS systems across two independent testing sets. Given their high performance, clinical utility, and illustrated interpretability, our models can be easily deployed in clinical settings to supplement existing EWS systems since they use the same data streams.