The representation of uncertainty in medical expert systems

The development of the rule-based expert system has provided important new techniques for the representation of knowledge. However, continued use of this representational scheme has highlighted some of its deficiencies. In particular, many within scientific and non-scientific fields attempting to use the rule-base design to describe natural phenomena often find it difficult to represent the complexities of the world as 'absolute' rules. For this reason, many investigators acknowledge the need to add an uncertainty mechanism to the rule-base construct. Such a facility would allow the quantification of accuracy or strength of association within individual rules Although agreement exists on the need for an uncertainty representation facility, the debate concerning the most appropriate methodology is far from resolved. The purpose of this paper is to provide a review and commentary on the current state of debate over the five most popular candidate uncertainty models: symbolic representation, MYCIN certainty factors, Bayesian, Dempster-Shafer and fuzzy set logic. The advantages and disadvantages of each uncertainty calculi will be presented and assessed with respect to their applicability to the medical expert systems domain.