Knowledge-based trend detection and diagnosis

This report presents a knowledge-based approach to diagnostic process monitoring. The cornerstone of this work is the representation and detection of multivariate trends in process data. The trend representation, called a trend template, denotes a time-varying pattern in multiple variables common to a diagnostic population. Each pattern contains representations for landmark events and a set of phases, each temporally uncertain. The phases are represented by a partially ordered set of temporal intervals. Bound to each interval are constraints on real-valued functions of measurable parameters. The constraints are low-order polynomial regression models, with either qualitative or quantitative coefficient estimates. A computer program called TrenDx diagnoses trends by matching process data to a set of competing trend templates within a clinical context. The matching score of a trend template hypothesis is based on the mean absolute percentage error between the regression models and the data. TrenDx not only maintains alternate hypotheses of different trends, but also optimizes over different chronologies within each trend description. Therefore TrenDx can report both what the most significant trend is and when events and phases take place within that trend. The report describes how TrenDx can be extended to complete an architecture for automated diagnostic process monitoring. A faulty trend is judged significant if over time it matches process data better than the expected trend. Significance of a faulty trend may trigger an alarm, switch the clinical context, or filter data for an intelligent display. TrenDx has been applied to diagnosis of trends in two medical domains. The program diagnoses trends in pediatric growth from heights, weights, bone ages, and sexual staging data. TrenDx also detects trends in intensive care unit patients from hemodynamic and respiratory data. The techniques of TrenDx are intended as general purpose, and may