MEDICAL DIAGNOSTIC TEST SEQUENCING AND OPTIMAL PROTOCOL DESIGN
Abstract
The purpose of this research is to examine diagnostic protocols as designed and optimization problems. Although our primary application concerns medical protocols, the methodology is sufficiently general to apply to industrial quality testing or repair procedures as well. Both from a theoretical and an applications perspective, protocol design is a new topic for mathematical optimization. Our method uses receiver operating characteristic (ROC) curves and information theory with differential calculus to determine optimal test decision levels. This level is the cutoff criteria to separate a positive from a negative test so as to minimize the cost of patient diagnosis and treatment. Cost is measured in any quantifiable unit such as dollars, cost per life year, etc. Optimality of a protocol is also dependent upon the sequence of tests in the protocol. A ranking procedure is used to sequence tests so they are dominant to all subsequent tests in the protocol. Dominance is based on diagnosis and treatment cost and upon the information gain provided by the test. Test sequencing and decision level optimization are interdependent problems and the algorithms established in this research allow simultaneous solution of both problems. Methods of handling test correlation are introduced and a procedure implemented to account for correlation between tests in our designs. The methodology of this thesis is used to design a protocol to minimize the cost per life year of diagnosing and treating patients with renal artery stenosis (RAS). Computer simulation is used to estimate some of the costs. Data from 607 patients seen at the Hypertension Center of Indiana University School of Medicine is used to formulate ROC curves for the diagnostic tests available for RAS screening. Results of the designed protocol indicate a potential savings of $188 per patient in test and treatment cost using the designed protocol instead of the current protocol, plus a slightly increased life expectancy resulting from a better diagnosis. This is equivalent to $17 per life year over an expected 12.8 years. The methodology is shown to produce a better solution than would be predicted by a random sample of complete enumeration of all test sequences.
Degree
Ph.D.
Subject Area
Operations research
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.