Using linear regression functions to abstract high-frequency data in medicine.

This paper investigates the problem of representing medical time series in linear piece-wise functions and proposes a novel algorithm to transform time-stamped numeric data into simple linear regression functions. We apply methods that involve the hat matrix leverage value and the studentized deleted residual to identify outliers, and a heuristic approach to remove them from the data sets. By distinguishing the breaking points from true outliers, we can efficiently break the data set with respect to the underlying patterns. Using a rough segmentation step, our approach avoids using the whole data set as input, and reduces space requirement. The experimental results indicate our method can achieve more accurate representation of the underlying patterns in data sets collected in the intensive care units efficiently.