![]() |
|||
![]()
|
![]() |
![]() |
![]() |
2. HISTORICAL ACCOUNTThe use of computers in medical decision-making began in the early 1960s with the implementation of programs that performed well-known statistical analysis. These programs focused on the diagnosis part of the consultation. They accepted a set of findings and selected one disease from a fixed set, using methods such as pattern recognition through discriminant functions, Bayesian decision theory, and decision-tree techniques. In the 1970s, medical expert systems were primarily concerned with performing diagnoses and making therapy recommendations (e.g., PIP, CASNET, MYCIN, and INTERNIST, CADUCEUS, PUFF, etc.). As early efforts, they were prototypes directed at two questions (Barr and Feigenbaum, 1982): What would make such a system acceptable to physician users? How can factual and judgmental knowledge be integrated? Despite the extensive work that has been done, none of these systems was in routine clinical use because of limitation of knowledge representation techniques and physician resistance, which will be discussed later. During the 1980s, we witnessed the proliferation of expert systems in medicine. They had a great impact on many areas of the medical field where knowledge provides the power for solving important medical problems. According to the survey conducted in 1992 (Durkin, 1994), the total number of expert systems uncovered in all fields was approximately 2500 in the 1980s. However, this number is estimated to represent only 20% of the total population -- 12,500 developed systems. Of these, 12% of the applications were in the field of medicine. A 1986 survey conducted by Waterman (1986) showed that an even higher percentage of the applications (30%) was in the field of medicine. These surveys showed that this field was very attractive to expert system developers in the 1980s. It is interesting to follow the growth rate of the expert systems in general. During the 1970s, when researchers were focusing on developing intelligent programming techniques, only a handful of systems were built. During the 1980s, the number of developed expert systems increased from 50 in 1985 to 2200 in 1988 (Harmon and Sawyer, 1990). The impressive growth rate of expert systems is an indicator of the acceptance of the technology by industry. These surveys also showed that expert systems are merging with the mainstream of information processing that was previously dominated by conventional data processors. We can attribute the large growth rate in developed systems in part to the new hardware and software technologies. During the 1970s, most expert systems were developed on powerful workstations, using languages such as LISP, PROLOG, and OPS. This placed the challenge of developing systems in the hands of the select few who could afford the platforms and had the patience to learn the complexities of the available languages. During the 1980s, we witnessed the proliferation of personal computers, and the introduction of easy-to-use expert system software development tools called "shells." A shell is a programming environment that contains all of the necessary utilities for both developing and running an expert system. During the last decade, large numbers of shells were marketed for use on personal computers. The opportunity to develop an expert system was now in the hands of many individuals, including physicians. According to the survey conducted by Kraut and Mann (1996), well-known expert system "shells" applied were ProMD (31%), HUGIN (19%), NEXPERT (12%), KAPPA (8%), and ClassicaD3 (8%). In 1990s, the complexity and volume of medical knowledge has increased continuously. As seen in the Appendix, a total of 233 medical expert systems were found from the following sources between 1992 and 1996: Medline database, Proceedings of the Medical Informatics Europe 96 (Brender et al., 1996), Proceedings of the 3rd World Congress on Expert Systems (Lee et al., 1996), and Proceedings of the 8th World Congress on Medical Informatics (Greenes et al., 1995). From the review of these articles, it appears that current medical knowledge at all levels of medical care was applied in developing these expert systems in order to achieve high-quality medical care as well as to reduce costs in medical care. Furthermore, a major change can be seen in the last few years. In the process of developing an expert system, the basic methodological problems like knowledge representation and inference mechanisms are no longer holding the spotlight. Rather, problems of introducing the systems in the clinical environment and questions of the application-oriented research are receiving the attention. Adequate user interfaces, a satisfactory cost/benefit ratio, and the adaptation of system functions to the requests of the clinical environment have become important research issues. 3. METHODOLOGIESSince the beginning of expert systems technology, knowledge acquisition and representation have long been considered the major constraint in the development of expert systems in the medical field. A majority of incidences of reported knowledge acquisition problems involved problems with the quality of knowledge elicited. Much of this was because doctors either tended to communicate shallow knowledge rather than the required deep knowledge structure, or they found it difficult to describe procedures and routines. In addition, knowledge acquisition from a standard clinical examination is also troublesome because patients responses are very subjective, and they may contradict themselves, sometimes repeatedly, when describing symptoms (Mouradian, 1990). Knowledge acquisition and representation is a kind of knowledge model that can be used to predict or explain behavior in the world. Thus, diagnosis is based on a causal explanation of what is happening to the patient, and therapy is based on predictions about how the disease process can be modified. Knowledge models in medicine are incomplete in that they are approximated and omit levels of detail (Clancey and Shortliff, 1984). One reason for this incompleteness is that there is no practical way to build systems that display more than a fraction of what any physician knows about the body and how it works. There is just too much knowledge, and we are still struggling to formalize even small portions of it. A second reason for leaving out levels of detail is that useful problem-solving performance can usually be produced even if we omit the pathophysiological knowledge about the disease. Among the knowledge models used in medicine are the following.
|
![]() |
|
Use of this site is subject certain Terms & Conditions. Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details. |