TrendFinder: Automated Detection of Alarmable Trends

Christine L. Tsien

As information technology continues to permeate all areas of our society, vast amounts of data are increasingly being collected for their potential utility. This is especially true of data-rich environments, such as airplane cockpits or hospital operating rooms. These environments share in common the presence of multiple sensors whose aim is to monitor the state of affairs in that environment over time. Data from sensors, however, can only be as useful as we know how to interpret them and are able to do so in a timely manner. If we have a good understanding of the relationship between monitored sensor values and the status of the monitored process, we could create knowledge-based systems, for example, to help interpret the volumes of data that would not otherwise be possible for a human observer to do. Often, however, such relationships are not well understood. How could these data be useful then? Skillful use of machine learning is one answer.

In this thesis, we present TrendFinder, a paradigm for discovering the knowledge needed to detect events, or trends, in numerical time series. Specifically, we present a suite of data collection, preprocessing, and analysis techniques that enable effective use of existing supervised machine learning methods on time-series data. We demonstrate how these techniques can be applied to the development of `intelligent' patient monitoring in the hospital intensive care unit (ICU), where currently as many as 86\% of bedside monitor alarms are false. First, we describe application of the TrendFinder paradigm to artifact detection in a neonatal ICU. Second, we describe TrendFinder's techniques applied to detection of trends indicative of `true alarm' situations in a medical ICU. Through illustration, we explore issues of data granularity and data compression, class labeling, multi-phase trend learning, multi-signal models, and principled time interval selection. We further introduce the idea of `post-model threshold refinement' for adapting a machine-learned model developed for one population to use on a different population. Overall, we demonstrate the feasibility and advantage of applying TrendFinder techniques to ICU monitoring, an area that is extremely important, challenging, and in need of improvement.

Submitted April 28, 2000 to the Department of Electrical Engineering and Computer Science in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Electrical Engineering and Computer Science.

Thesis Supervisor: Peter Szolovits, Professor of Electrical Engineering and Computer Science


Last modified: Mon Jun 19 08:36:08 EDT 2000
Jon Doyle <doyle@mit.edu>