Control Issues in Classificatory Diagnosis

Abstract

A good part of medical diagn osis can be modeled as classification problem solver producing a "differential", working in conjunction with an abductive component that perfor ms differential diagnosis by synthesizing a "best" composite hypoth esis out of the hypoth eses in the differential list . Classification problem solving itself can be viewed as having a control component whi ch selects hyp ot heses to consider, and a decision component associated with each selected hypothe sis. In this paper we st udy the family of control regimes that are useful in classificatory problem solving. We start with MDX, a cla ssification system organized as a hierar chical collection of hyp ot hesis specialists, critique it s control behavior , and by considering a set of sit ua tions involving multiple di seases, show how elements can be added to the cont rol regime in a modular way to handle a large variety of s ituati ons.