Abstract
This paper presents a two-leveled hybrid system for medical diagnosis. The first level is responsible for hypotheses' selection and uses an original association-based reasoning scheme (shallow knowledge), which measures the distance between the observations and the prototypical models of diseases, using fuzzy decision functions. This first module is efficient but not very precise and hardly transparent; the key of the representation it uses is to understand symptoms which occur within a disease's definition as fuzzy criteria, and to aggregate these criteria in a unique complex decision function which models the disease as a whole. The reduced problem made up of the hypotheses selected in the first step is passed through a refining and discriminating process, which was inspired from direct argumentation systems (the latter being a relatively new approach to diagnosis problems). This part is meant to provide explanation facilities for the results and to remove the contradictions generated by the previous step, if any. Original definitions are suggested for argument (a monotonic structure replacing abductive explanation) and attack (representing the defeasibility/non-monotonicity of reasoning). It is more efficient to reduce non-monotonicity to an attack relation between arguments, than any of the approaches used by hypothetical-deductive abduction. A brief theoretical analysis of the model within Ginsberg's unified framework of multivalued logic is finally presented.