AAAI 1992 Spring Symposium Series Reports

Unlike previous Artificial Intelligence in Medicine (AIM) meetings, the AIM Symposium was organized around a single source of medical knowledge, the 1991 article “Graft-Versus-Host Disease (GVHD)” by Dr. James L. M. Ferrara and Dr. H. Joachim Deeg (New England Journal of Medicine 324:667). Presenters were also given three clinical vignettes in GVHD to use as examples in their talks or poster presentations. The symposium was significantly more focused because of the use of a common knowledge source. Provocateurs had the task of asking difficult questions to both the presenters and the audience. Also, because most participants had carefully examined the article from the perspective of their own research, the depth of questions and discussion was vigorous and extensive. The symposium was dominated by two themes: knowledge sharing and reuse and temporal reasoning. Additional submissions on diagnostic reasoning and knowledge representation methods were also represented. The symposium began with an overview talk by Peter Szolovits entitled “AI in Medicine: Past and Future.” In this talk, Szolovits reminded the group that AIM has continued to make steady progress on difficult issues such as multiple diagnoses, probabilistic reasoning, and (simple) temporal reasoning. He also noted that like other AI application disciplines, AIM has moved away from a focus on core AI issues such as planning and learning. Although it remains a rich area for research by AI specialists, AIM must now also address a broad and difficult interdisciplinary mix of topics, including information systems, data analysis, man-machine interfaces, pathophysiology, and genetics. Theme one of the symposium was knowledge acquisition and knowledge reuse. Talks by Mario Stefanelli (Universita di Pavia), Jeff Bradshaw (Boeing Computer Services), Samson Tu (Stanford University), Nasir Amra (The Ohio State University), and William Punch (Michigan State University) focused on various methodologies for acquiring and reusing GVHD knowledge. Ramesh Patil (USC/ Information Sciences Institute [USC/ISI]) presented the Knowledge Sharing Initiative, and Michael Kahn (Washington University) contrasted this work with the American Society for Testing Materials 31.15/Arden Syntax work. The presentations and discussion on this theme illustrated that a significant amount of progress has been achieved in building tools that allow for the rapid representation and evaluation of qualitative simulation and rule-based models in the AIM community. Theme two was temporal reasoning. Tom Russ (USC/ISI), A. Mete Kabakcioglu (University of Miami), Yuval Shahar (Stanford), Issac Kohane (Harvard University), Ira Haimowitz (Massachusetts Institute of Technology [MIT]), Barbara Heller (University of Bielefeld), and Gregory Provan (University of Pennsylvania) described markedly different approaches to encoding and reasoning about the dynamic aspects of GVHD. Provocateur Jim Hunter (University of Aberdeen) closely critiqued the various approaches to temporal reasoning. Hunter also noticed that most of the temporal ontologies were simple point-based methods that did not encode temporal uncertainty (a point also made by Szolovits in his opening talk). He also noted that much work in the more traditional AI fields was relevant to AIM problems but that this research did not appear to be used by AIM researchers. The remaining talks were dominated by diagnostic and knowledge representation methodologies: criteriatable methodology by May Cheh (National Library of Medicine) and Kent Spackman (Oregon Health Sciences), qualitative models by Serdar Uckun (Vanderbilt University), and a hierarchical knowledge representation methodology by H. Mannebach (North Rhine-Westphalia Heart Center). Researchers also presented a wide variety of methodologies during the poster session: Steve Cousins (Washington University) on the use of query networks, Peter Hucklenbroich (GSF-MEDIS-Institut) on the KLINC knowledge representation model, Donna Hudson (University of California at San Francisco) on a hybrid expert system model, Yeona Jang (MIT) on a causal graph diagnostic model, M. Magues (Institut European de Telemedecine) on a telemedicine assistant, John Weiner (University of Southern California) on a concept structure for representing consensus, and Jeremy Wertheimer (MIT) on the representing and the reasoning about molecular physiology. The symposium was successful because all the participants had invested a significant amount of work in understanding a specific medical disease. It was unnecessary for each speaker to introduce his(her) specific problem domain. The examples used in the talks and posters were substantive because the presenters could assume that the audience had a similar knowledge of the example disease. Because the content of the symposium was greatly improved by this structure, it is hoped that future AIM meetings will adopt a similar model, perhaps selecting other articles that address segments of the AIM community that were