Metareasoning and meta-level learning in a hybrid knowledge-based architecture

Multiple representations and multi-modal reasoning in medical diagnostic systems

The paper examines the motivations for developing medical diagnostic systems exploiting multiple representations and multi-modal reasoning. The analysis is carried on by revisiting the architectural choices of the CHECK system (developed in late 1980s) which combined heuristic and causal knowledge. The results in the theory of diagnosis and in model-based reasoning (MBR) obtained in early 1990s are used for providing a formal characterization of the notion of diagnosis and of the reasoning mechanisms used in CHECK. The paper addresses also the problem of replacing heuristic knowledge provided by human experts with operational knowledge automatically derived from the deep model. In particular, the pros and cons of knowledge compilation and of the integration of case-based reasoning (CBR) with MBR are discussed by summarizing the experience gained in developing AID and ADAPtER. The problem of using an explicit representation of time in diagnostic systems is analyzed and recent work on the different characterizations of diagnosis arising when the temporal dimension is considered is reported. Finally, the implications of the results obtained in MBR and in temporal reasoning on the future of medical diagnostic systems are briefly discussed.