High performance for expert systems: II. A system for leukaemia diagnosis

A system for diagnosing leukaemia subtypes from monoclonal antibody test results has been programmed directly from the expertise of one immunologist. It gives an acceptable conclusion for 400 past cases, with summaries of its reasoning and suggestions for further tests. Its development included three major components: elicitation of the basic items of knowledge, formulation of a coordinated overall scheme, and representation in a computer language. All three were interdependent and required many cycles of trial and error, involving further discussions with the expert at all stages. By far the most difficult component was the formulation of a comprehensive specification for the problems of the domain, and this was only achieved when the knowledge engineer had gained a fair understanding of the application. The system is a tree-structured logic program written in PROLOG. This format facilitated experimentation with different knowledge representations, control of the interrelationships between rules and the elimination of errors. The problems of quality control and maintenance are of paramount importance for high-performance systems.