Diagnostic process monitoring with temporally uncertain models

Abstract

This thesis develops a real-time trend detection and monitoring system based on previous work by Haimowitz, Le, and DeSouza [3, 5, 2]. The monitor they designed, TrenDx, used trend templates in which the temporal points where data patterns change are variable with respect to the actual process data. This thesis uses similar models to construct a monitoring system that is able to run in real time, based on a continuous, linearly segmented process data input stream. The instantiation of temporally significant template points against the process data is determined through a simulated annealing algorithm. The rankings of competing hypotheses in the monitor set is based on the distance of these template points from their expected temporal values, along with the area between the process data measurements and the value constraints placed on those parameters. The feasibility of the real-time monitor was evaluated in the domain of pediatric growth, particularly in comparison to previous versions of TrenDx, using an expert gold standard of the diagnoses of pediatric endocrinologists. Real-time TrenDx shows promise in its monitoring abilities and should be evaluated in other domains which are more suited to its continuous data stream input model.