In modern intensive care physiological variables of the critically ill can be reported online by clinical information systems. Intelligent alarm systems are needed for a suitable bedside decision support. The existing alarm systems based on fixed treshholds produce a great number of false alarms, as the change of a variable over time very often is more informative than one pathological value at a particular time point. What is really needed is a classification between the most important kinds of states of physiological time series. We aim at distinguishing between the occurence of outliers, level changes, or trends for a proper classification of states. As there are various approaches to modelling time-dependent data and also several methodologies for pattern detection in time series it is interesting to compare and discuss the different possibilities w.r.t. their appropriateness in the online monitoring situation. This is done here by means of a comparative case-study.
Online monitoring time series analysis state classification change point detection ARIMA models phase space models dynamic linear models