Signal and variability extraction for online monitoring in intensive care

Abstract

This thesis proposes new methods for real-time signal and variability extraction, presents derivations of their robustness properties and discusses their value for practical applications to physiological time series. Although the proposed techniques are developed against the background of online monitoring in intensive care, they are also applicable to any other kind of time series. For Repeated Median regression on an equidistant grid, the distribution of the position and number of zero residuals is investigated, and the correlation structure between the residual signs is examined. For online signal extraction, an adaptive filter is proposed which essentially relies on a goodness-of-fit test based on residual signs from Repeated Median regression. After deriving suitable settings for this filter in the univariate case from a simulation study, the procedure is extended for application to multivariate time series. For online variability extraction, three approaches to scale estimation are considered. The robustness properties of the newly proposed regression-free and model-free techniques are derived, and the different approaches are compared via an extensive simulation study.