This statement is, alas, rather old and somewhat out of date. My apologies.Since the early 1970's, the thrust of Artificial Intelligence (AI) research has shifted from the search for very powerful general problem solving techniques to the design of complex knowledge representation formalisms in which large amounts of rather specific task-oriented knowledge may be encoded and effectively accessed. This shift is motivated by the observation that human expert performance appears to be based on knowing just what to do in many, many situations rather than on always being able to work out the appropriate solutions from first principles. I am working on the development and the application of such representation techniques and their associated reasoning routines, to build what have come to be called "knowledge based systems" or "expert systems." The application area I have chosen is medicine. I work with a group of physicians to identify and program simulations of the strategies which expert doctors bring to bear on difficult medical problems. We also develop engineering methods that can augment the capabilities of human reasoning. Our aims are threefold: (1) to use this rich, realistic problem domain as a challenge to suggest and test out new AI ideas, (2) to clarify the nature of medical knowledge and reasoning, and (3) to build computer programs that embody medical expertise and can serve as consultants, as error monitors and as educational tools to improve health care.
Because of vastly improved and cheaper computation and communication technologies, the 1990's provide great opportunities for engineering innovation in the use of computerized clinical data. We have begun one project to make the clinical database of the Children's Hospital available via the World Wide Web (WWW), subject to careful privacy and security provisions. Contrary to most past assumptions that medical data is somehow unique, we are exploring the hypothesis that its collection and dissemination raises issues also addressed (and often solved) in other commercial sectors. Thus, we expect to take advantage of WWW servers and browsers to provide multi-platform access available worldwide, with standard encryption and authentication mechanisms. Difficult issues in this project include a common representation for medical data that may be stored in heterogeneous forms in various databases, and customization of interfaces to best support the needs of individual users. In another new project, Guardian Angel, we plan to build life-long comprehensive personal medical information systems that focus on the medical information needs of the individual rather than doctors or institutions. Such systems involve computational processes that are active for the entire lifetime of their subject, and are responsible for collecting, storing, and disseminating all medically-relevant data about a patient, interpreting those data to the patient and educating him or her to their implications, "sanity checking" medical plans against what is already known about the patient, and representing the patient's preferences in scheduling, treatment planning, etc. We are beginning to explore these ideas in some specific medical domains.
During the late 1970's and early 1980's, my group led the medical AI community in moving away from simplistic pattern-matching approaches to diagnostic reasoning, toward the use of much richer models that incorporate reasoning about the physiological mechanisms of disease, quantitative considerations, relative likelihoods, models of temporal evolution of disease, and methods of combining simple "it is what it looks like" with deeper model-based reasoning. Much of the AI field in general has now adopted the pure model-based view, but unfortunately at the cost of returning to the first-principles methods whose intractability originally spawned the knowledge-based movement. I believe that the challenge continues to be how to find a powerful but efficient combination of the two.
Within the past decade, we have made a great deal of progress on various aspects of this overall problem. We have developed an initial model for combining deep and shallow reasoning, we have implemented several different reasoning programs that exploit various aspects of temporal reasoning, we have made some advances on the encoding of uncertainty, and we have developed a much better understanding of what it means to reason qualitatively about complex systems. With the recent advent of data-rich hospital information systems, we are now also engaged in the practical application of many of these ideas with collaborators at Boston area hospitals.
Our group has produced an outstanding group of doctoral graduates in the past few years, including Mike Wellman, Elisha Sacks, Phyllis Koton, Alex Yeh, Tom Russ, Tom Wu, Yeona Jang, Ira Haimowitz and Tze-Yun Leong. We continue to have excellent new students, and are engaged in a wide range of research areas. In addition to basic medical reasoning and its applications, we are developing efficient means to evaluate probabilistic models and applying them in clinical genetic counseling and molecular genetics. I am also pursuing a more theoretical study demonstrating the equivalence of the known exact methods for Bayesian calculation with colleagues from my sabbatical at Stanford; I have been arguing, with Davis and Shrobe, for an expansion of the agenda of knowledge representation work; and three years ago I co-chaired the program committee for the national AI conference, hoping to influence the review process toward broader themes. Steve Pauker and I were honored by the selection of our 1978 AI Journal paper as one of the 25 most often referenced papers in that journal, and we wrote a follow-up article. I was also recently elected a Fellow of the American Association for Artificial Intelligence. With colleagues at Harvard and Tufts medical schools, we were awarded a major training grant in medical informatics, which will allow us to train a new generation of medical AI specialists.