A Bayesian computer-based approach to the physician's use of the clinical research literature

To date, automated statistical methods used to help physicians use the clinical research literature for making clinical decisions have been limited in the degree to which they can represent methodological and domain concepts that are crucial to the physician who must take clinical action. In this dissertation, I consider the thesis that Bayesian decision theory can provide the foundation for a computer-based environment that helps physicians to use the research literature. On the basis of a knowledge-level analysis of this problem, I argue for the use of Bayesian statistics over classical statistics. To show that this new paradigm can be implemented in a functioning computer system, I have developed a prototype system, called scTHOMAS, that enables a physician to read a research report, to incorporate her domain knowledge and methodological concerns, and to evaluate their impact on the clinical significance of the conclusion. The system effectively automates the Confidence Profile Method of Eddy, Hasselblad, and Shachter $\sim$ (1991). scTHOMAS operates in the domain of randomized clinical trials that compare the effects of different drugs on a patients' survival. To incorporate any methodological concern, scTHOMAS (1) requires a statistical submodel for the concern, and (2) requires a visual metaphor though which the physician can communicate the particular concern. scTHOMAS contains submodels for the methodological concerns of loss to followup, withdrawal, noncompliance, crossing-over, and measurement unreliability. The system uses the visual metaphor of the patient-flow diagram for physician input. In the course of each consultation, the user implicitly constructs a statistical model. Statistical models are represented as hierarchical, typed influence diagrams, a structure that limits the interactions among parameters in a statistical model. Prespecified construction steps dictate how the primitive methodological submodels are pieced together. A metadata-state diagram, containing basic methodological knowledge assessed from a statistical expert and from the methodological literature, limits the sequence of construction steps the user is allowed. This dissertation puts on the medical-informatics agenda the question of how physicians should act on the basis of research data, and suggests novel methods for storing, using, and retrieving the contents of the biomedical research literature.