Artificial Intelligence in Medicine

Edited by Peter Szolovits

(AAAS Selected Symposia Series, Volume 51)

1982

This book introduces the field of artificial intelligence in medicine, a new research area that combines sophisticated representational and computing techniques with the insights of expert physicians to produce tools for improving health care. An introductory chapter describes the historical and technical foundations of the work and provides an overview of the current state of the art and research directions. The authors then describe five prototype computer programs that tackle difficult clinical problems in a manner similar to that of an expert physician. The programs presented are INTERNIST, a diagnostic aid that combines a large database of disease/manifestation associations with techniques for problem formulation; EXPERT and the Glaucoma Program which use physiological models for the diagnosis and treatment of eye disease; MYCIN, a rule-based program for diagnosis and therapy selection for infectious diseases; the Digitalis Therapy Advisor, which aids the physician in prescribing the right dose of the drug digitalis and also explains its actions; and ABEL, a program that uses multi-level pathophysiologic models for diagnosis of acid-base and electrolyte disorders.

2000 Note

This book has been out of print since the early 1990's, though it is still often available through book search services, including the usual on-line ones. The original book was relatively popular for a symposium book; it was re-published in paperback and went through several printings. From the vantage point of nearly twenty years after its publication, I believe that many of the ideas in the chapters are still vibrant. Sophisticated modeling in artificial intelligence approaches to medical reasoning have to a significant extent been supplanted by attempts to exploit knowledge implicit in large clinical datasets via machine learning techniques. At the same time, medical record systems have moved toward routine adoption so slowly that the authors would have been shocked in 1982 to discover that many of the ideas we described are still immensely difficult to apply in practice because the data they rely on are not normally available in machine-readable form.

I have reconstructed this volume in HTML and made it available on the Web in the hope that it will inspire new researchers to learn about some of the elegant older work and to take up the challenges not yet met.

Conventions for the appearance of books as Web documents have not yet been firmly established. In the present attempt, I have chosen to encode each chapter as a single document (with the exception of a long transcript in Chapter 5, which is linked as a separate page). Figures appear in-line, at low resolution. Where this is insufficient to allow the reader to interpret details, these images are hyperlinked to very large 300dpi versions. Layout of pages in the original is, of course, completely lost in this form. I have refrained from editing the text except to correct a few obvious typographic and editorial errors in the original. It must be read, therefore, as a work from the early 1980's, without benefit of our current knowledge of its future. I hope that the style of this work will be acceptable, and that its content will intrigue.

About the Editor and Authors

Peter Szolovits is an associate professor in the Department of Electrical Engineering and Computer Science at the Massachusetts Institute of Technology. A specialist in artificial intelligence (Al) with an emphasis on medical applications, he is currently concerned with fundamental issues of representation and reasoning, including protocol analysis to discover how clinicians reason about probability and causality, and with programs which model human expert performance in some areas of medical diagnosis and care.

Randall Davis is an assistant professor in the Department of Electrical Engineering and Computer Science at Massachusetts Institute of Technology. His research interests are knowledge-based systems, knowledge representation, problem-solving methods, and use of artificial intelligence in medicine. He is the author of Knowledge-based Systems in Artificial Intelligence (with D. B. Lenat; McGraw-Hill, 1982).

Casimir A. Kulikowski is professor of computer science and associate director of the Laboratory for Computer Science Research at Rutgers University; since 1972 he has also been a senior investigator and associate director of the Rutgers Research Resource on Computers in Biomedicine. His research is in the fields of artificial intelligence and pattern recognition, with emphasis on expert systems and their applications, and he has directed several collaborative projects for the development of expert medical consultation systems.

William J. Long is on the research staff at the Laboratory for Computer Science, Massachusetts Institute of Technology. Currently he is working on development and evaluation of the Digitalis Therapy Advisor and on the design of AI programs for heart failure diagnosis and therapy, and he has written on the criteria for computer-generated therapy advice in a clinical domain.

Ramesh S. Patil, a research assistant at the Laboratory for Computer Science, Massachusetts Institute of Technology, has done research on the design of a program for the causal understanding and diagnosis of patient illness in electrolyte and acid-base disturbances.

Harry E. Pople, Jr., is an associate professor of business and co-director of the Decision Systems Laboratory at the University of Pittsburgh. A computer scientist, he has specialized in artificial intelligence with emphasis on clinical decision making, problem formulation, and problem solving.

William B. Schwartz is Vannevar Bush University Professor and professor of medicine at Tufts University and associate member of the Laboratory for Computer Science at Massachusetts Institute of Technology. A physician by training, he has published extensively on renal function and electrolyte and acid-base equilibria in humans. He is also currently involved in development of medical applications of computers, including their use for decision analysis and the simulation of clinical judgment.

Sholom M. Weiss, an associate research professor of computer science at Rutgers University, has worked on developing computer-based consultation systems in medicine. His current research interests are the development of general systems for designing expert models and the application of expert systems in medicine, laboratory instrumentation, and oil exploration.

Preface

This book presents an overview of the major research efforts in the United States in the application of artificial intelligence techniques to medical decision making. It has grown out of a Symposium on Artificial Intelligence in Medicine presented at the 1979 Annual Meeting of the American Association for the Advancement of Science. The chapters in this volume are mainly revisions of presentations at the Symposium, augmented to bring the work more up to date and to address questions arising from the audience and discussions at the meeting.

The Artificial Intelligence in Medicine (AIM) field emerged in the early 1970's in response to several simultaneous needs, opportunities and interests. An increased demand for high-quality medical services coupled with the explosive growth of medical knowledge has led to the suggestion that computer programs could be used to assist physicians and other health care providers in discharging their clinical roles in diagnosis, therapy and prognosis. At the same time, computer science techniques, especially those of the artificial intelligence field, began to reach a maturity with which they could be applied to representing and reasoning about complex, "real world" problems like those arising in medicine. Investigators trained on both the computational and the medical side of these concerns began to develop mutual interests and approaches, and to form coherent collaborative research efforts which have now produced the programs reported on in this volume.

Modern medicine has become technically complex, the standards set for it are very high, conceptual and practical advances are rapid, yet the cognitive capabilities of physicians are pretty much fixed. As more and more data become available to the practicing doctor, as more becomes known about the processes of disease and possible interventions available to alter them, practitioners are called on to know more, to reason better, and to achieve better outcomes for their patients. But how can the physician learn, retain and apply this ever-expanding body of knowledge?

Improvements in medical education are a constantly sought but limited means of meeting this challenge. The creation within our medical centers of specialty and sub-specialty services, with experts serving as shared consultants to the rest of the community and training programs to create more experts has been the most important institutional response. However, such a structure also creates major differences among users of the health care system. The general practitioner (or today's "family practice specialist") can no longer embody both a broad and deep expertise in all of medicine. To get the best of care, the patient may need the advice of several specialists over the course of time. Although major medical centers can and do provide such an adequate spectrum of services, those economically disadvantaged or remotely located cannot or do not receive such an intensive concentration of assistance. This problem is especially acute in underdeveloped countries, where even the supply of primary physicians, nurses and technicians is very limited.

If the expertise of consultants can be captured in the form of computer programs which provide advice to less-expert physicians or other health-care providers, then any practitioner could call on that expertise whenever a patient's case suggested the need for careful thought about some aspect of the illness or therapy. The continued astonishing increase in power and decrease in price of computers will put a large (by today's standards) computer within reach of every physician's desk within a few years. The opportunity is there to improve the health-care system by improving each physician's ability to utilize the best available knowledge and the best ways of analyzing medical problems, as encoded in easily-duplicated and updated computer programs.

Implementing this vision is by no means simple. Even technical perfection is no guarantee that a technology will be adopted and used in the ways its proponents envision. What social, ethical, legal, professional, or other reasons may arise to step in the way? And of course technical success is also still more a promise than a reality. The programs described here have made a good start, and have demonstrated that the vision is feasible. Several of them, with needed additional development work, could probably already serve to improve medical care in their areas of competence. The programs have also identified numerous rich problems for artificial intelligence research, and have put their developers at the forefront of research in At applications.

In this collection of papers, the authors invite you, our readers, to appreciate the excitement of the challenges we face, the accomplishments we have achieved so far, and the progress we plan. The reports included here are from the major research groups in the U.S. No attempt has been made to be exhaustive in coverage, though the best-known programs in the field are represented. Fortunately, the related work of our colleagues and students at our own universities and at other domestic and foreign universities and research centers is too voluminous to include in this work -- our field is growing, attracting brilliant young computer scientists and physicians, and finding institutional and financial support.

As editor, I would naturally like to thank all the contributors, whose preparation for the original Symposium and later preparation and revision of these chapters has undoubtedly seemed a never-ending task. I thank Dr. William B. Schwartz, my longtime collaborator, for talking me into editing this volume, and Ms. Joellen Fritsche, who gave editorial advice and bore patient witness for the AAAS during the long birth of this book. I thank my former secretary, Ms. Anne Schmitt, for help with typing and editing the manuscript and drawing illustrations, and my students for helping to read, analyze, and suggest corrections. My wife, Dianne Foster, has been kind enough to match my late nights of editing with almost equally late nights of her own legal studies. All the research reported on here has been touched by the supporting hand of the National Institutes of Health's Division of Research Resources and the National Library of Medicine, two institutions whose leaders have recognized the promise ofthe field and devoted themselves to its progress.

This book was prepared using computerized text editing, formatting and typesetting systems at MIT's Laboratory for Computer Science. The EMACS text editor, developed by Richard Stallman, has been an excellent tool supporting the incremental editing of countless drafts of the manuscripts. The book was formatted and electronically typeset by Donald Knuth's TeX. Camera-ready copy was produced by our Dover printer, created at Xerox Palo Alto Research Center. During development of the book, as well as in the course of our research, the authors were in contact by the electronic mail facilities of the ARPA network, which allowed us to shuttle manuscript fragments around the country and gave us virtually instant mail. Computer-aided preparation of a book. an experience likely to be shared by growing numbers of authors, is somewhat of a mixed blessing--the authors gain great control over the appearance of the final product, but at the cost of considerably more work that traditionally belongs to the publisher. After this description, it should be more than obvious that flaws in the book belong squarely to the editor.

Publication Data for "Artificial Intelligence in Medicine"

Published by Westview Press, Inc.
5500 Central Avenue, Boulder, Colorado
for the
American Association for the Advancement of Science
1776 Massachusetts Avenue, N.W., Washington, D.C.

This book is based on a symposium which was held at the 1979 AAAS National Annual Meeting in Houston, Texas, January 3-8. The symposium was sponsored by AAAS Section N (Medical Sciences).

All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher.

Published in 1982 in the United States of America by
Westview Press, Inc.
5500 Central Avenue
Boulder, Colorado 80301
Frederick A. Praeger, President and Publisher

About the Series

The AAAS Selected Symposia Series was begun in 1977 to provide a means for more permanently recording and more widely disseminating some of the valuable material which is discussed at the AAAS National Meetings. The volumes in this Series are based on symposia held at the Meetings which address topics of current and continuing significance, both within and among the sciences, and in the areas in which science and technology impact on public p oh cy. The Series format is designed to provide for rapid dissemination of information, so the papers are reproduced directly from the camera-copy submitted by the authors. The papers are organized and edited by the symposium arrangers who then become the editors of the various volumes. Most papers published in this Series are original contributions which have not been previously published, although in some cases additional papers from other sources have been added by an editor to provide a more comprehensive view of a particular topic. Symposia may be reports of new research or reviews of established work, particularly work of an interdisciplinary nature, since the AAAS Annual Meetings typically embrace the full range of the sciences and their societal implications.