Greg Strylewicz

Clinical laboratories provide a critical service to the health care and well-being of the world's population. Estimates suggest that the clinical laboratory influences some 70 percent of health-care decisions, but requires only about 4 percent of the health-care expenditures. Given an estimated 7 billion laboratory tests per year in the United States, about 1% of the results, or 70 million laboratory errors annually, are erroneous with an estimated 6%, of those errors causing harm to the patient. Laboratory errors harm millions of patients each year and laboratory experts spend countless hours reviewing billions of laboratory results each year in the search for these rare errors. Autoverification systems, automated programs used to check laboratory results for errors, can save laboratories countless hours and be more accurate than laboratory experts, but the current generation of rule-based systems is not appropriate for the clinical laboratory domain due to its inherent uncertainty. This research demonstrates that a novel approach using a synthetic error generation system to create training datasets for a conditional Gaussian Bayesian network produces an autoverification system superior to ones trained using standard methods and superior to laboratory experts. Unlike standard approaches that require an expensive and time-consuming expert annotation process to create training datasets, the synthetic error generation method uses results that were reviewed normally.

By creating synthetic datasets, the synthetic error generation process creates customized training datasets, which maximize the Bayesian network's performance in detecting errors. In this dissertation, we review the clinical laboratory process and the many sources of errors in clinical laboratory results, Bayesian networks, and the class imbalance problem. Next, we elucidate the performance characteristics of the synthetic error generation process, which is followed by a comparison between our novel method and standard approaches to the class imbalance problem. Finally, we compare the results of a synthetic error autoverification system against laboratory experts in the identification of errors.