Extreme Value Theory for Novelty Detection in Vital Sign Monitoring

Studies indicate that a large number of in-hospital cardiac events could be prevented, provided that the preceding physiological abnormalities are identified and acted upon. Continuous patient monitoring is not a task that can reasonably be assigned to the nursing staff, but one that could benefit from the use of an automated system. With this in mind, an early warning system based on a data-driven novelty detection scheme has been designed by Tarassenko et al. [2005]. Vital sign data collected from clinical trials are fused to create a multi-patient ‘model of normality’ in the form of a multivariate probability density function. A novelty threshold is set in the probability space and used to discriminate between normal and abnormal test data. Performances of this methodology on clinically reviewed data are reported in Hann [2008]. In this investigation, we attempt to extend the research in two directions. First, we introduce a principled method for assigning novelty scores and setting novelty thresholds based on an extension of Extreme Value Theory, theory which focuses on describing the tails of univariate distributions. For the purpose of this work, we redefine ‘extrema’ to include all regions of low probability and propose a methodology for the study of multivariate distributions. We show that our novel scoring system is more stable and easier to interpret than the original one, and requires very little postprocessing. Second, we use the first 24 hours of a patient’s vital sign data to build a patient-specific model. Using our scoring system, we compare its results with those from the multi-patient model. On both our test and control group, we show that if patient’s vital signs are indeed normal over the training period, the patient-specific score is in turn less volatile and responsible for fewer false alarms.