Reference: Lehmann, H. P. & Shortliffe, E. H. THOMAS: Building Bayesian Statistical Expert Systems to Aid in Clinical Decision Making. 1991.
Abstract: Previous knowledge-based systems for statistical analysis separate the numeric knowledge in the data analysis from the heuristic knowledge in using the results of the analysis. In contrast, a Bayesian framework for building biostatistical expert systems allows for the integration of the data-analytic and decision-making tasks. The architecture of such a framework entails enabling the system (1) to make its recommendations on decision-analytic grounds, (2) to update a statistical model on the basis of data from the study and the userUs prior beliefs, and (3) to construct those models dynamically. This architecture permits the knowledge engineer to represent a variety of types of statistical and domain knowledge, including methodological knowledge. Constructing such systems requires that the knowledge engineer reinterpret traditional statistical concerns, such as by replacing the notion of statistical significance with that of a pragmatic clinical threshold. The user of such a system can interact with the system at the level of general methodological concerns, rather than at the level of statistical details. We demonstrate these issues with a prototype system called THOMAS which helps a physician reader to interpret the results of a published randomized clinical trial for clinical decision making.