Bayesian Optimisation of Gaussian Processes for Identifying the Deteriorating Patient

Abstract

Patient deterioration in the hospital ward is typically preceded by several hours of deranged physiology, as measured by the patient's vital signs. Estimation of the expected trajectory of a patient's future vital signs can allow us determine the degree of risk of physiological deterioration for that patient. Gaussian processes (GPs) offer a principled means of estimating vital-sign trajectories within a probabilistic framework. The automated estimation of GP parameters in this setting is difficult, due to the (often substantial) variation in physiology between patients, and also due to any changes in physiology that may occur for individual patients. Population-based techniques for fitting models to patient vital-sign data may be inferior compared with patient-specific approaches. We here propose the use of Bayesian optimisation to learn patient-specific models that are effective for estimating future physiological data, based on previously-observed data for the individual patient. We show how patient-specific values of GP hyperparameters may be learned using Bayesian optimisation, based on data observed during the first day of a patient's stay on an acute ward. We then demonstrate the benefit of using such methods in terms of forecasting accuracy for monitoring the patient during their subsequent two days on the ward.