Bayesian Condition Monitoring in Neonatal Intensive Care
The observed physiological dynamics of an infant receiving intensive care contain a great deal of information about factors which cannot be examined directly, including the state of health of the infant and the operation of the monitoring equipment. This type of data tends to contain both common, recognisable patterns (e.g. as caused by certain clinical operations or artifacts) and some which are rare and harder to interpret. The problem of identifying the presence of these patterns using prior knowledge is clinically significant, and one which is naturally described in terms of statistical machine learning. In this thesis I develop probabilistic dynamical models which are capable of making useful inferences from neonatal intensive care unit monitoring data. The Factorial Switching Kalman Filter (FSKF) in particular is adopted as a suitable framework for monitoring the condition of an infant. The main contributions are as follows: (1) the application of the FSKF for inferring common factors in physiological monitoring data, which includes finding parameterisations of linear dynamical models to represent common physiological and artifactual conditions, and adapting parameter estimation and inference techniques for the purpose; (2) the formulation of a model for novel physiological dynamics, used to infer the times in which something is happening which is not described by any of the known patterns. EM updates are derived for the latter model in order to estimate parameters. Experimental results are given which show the developed methods to be effective on genuine monitoring data.