An Epistemology for Clinically Significant Trends

We have written a computer program called TrenDx for automated trend detection during process monitoring. The program uses a representation called trend templates that define disorders as typical patterns of relevant variables. These patterns consist of a partially ordered set of temporal intervals with uncertain endpoints. Attached to each temporal interval are value constraints on real-valued functions of measurable parameters. As TrenDx receives measured data of the monitored process, the program creates hypotheses of how the process has varied over time. We introduce the importance of a distinct trend representation in knowledge-based systems. Then we demonstrate how trend templates may represent trends that occur at fixed times or at unknown times, and their utility for domains that are quantitalively both poorly and well understood. Finally we present experimental results of TrenDx diagnosing pediatric growth disorders from heights, weights, bone ages, and pubertal data of twenty patients seen at Boston Children's Hospital.