With the increased use of EHRs, it is important to develop new approaches to abstracting patient data to provide clinical decision support. In particular, we look at modeling vital signs. Such data are typically represented as a time series. Our goal is to label these time series with one or more integer labels that are equivalent to an expert-supplied evaluation of the quality, volatility, or level of alarm that should be associated with the time series.
Unfortunately, we do not expect classical shape- or pattern-based time-series classification methodologies to perform well, as, in practice, human experts assessing vital sign data seem to be evaluating the statistical properties of the series, rather than searching for specific patterns. We propose a novel, Bayesian statistical model, the AR-OR (AutoRegressive- Ordinal Regression) model, for labeling time series data and compare our approach to traditional time series classification methods.