![]() |
|||
![]()
|
![]() |
![]() |
![]() |
4.5. EEG ANALYSIS SYSTEMThis system was developed for the automated detection of spikes and sharp waves in the EEG for the detection of epileptiform activity (Davey et al., 1989). The system consists of two distinct stages. The first is a feature extractor, written in FORTRAN, which uses spike detection algorithms to produce a list of all spike-like occurrences in the EEG. The second stage, written in "shell," OPS5, reads the list and uses rules incorporating knowledge elicited from an EEGer to confirm or exclude each of the possible spikes. Information such as the time of occurrence, polarity, and channel relationship is used in this process. 4.6. OTHER EXPERT SYSTEMS DEVELOPED IN THE 1990SExperts systems were applied to the following medical fields in the 1990s: surgery (acute abdominal pain), microbiology (borreliosis), radiology (thoracic diseases), pharmacology, neurosurgery (meningitis), anesthetics and intensive care, pediatrics (metabolic diseases), neurology (cerebrovascular diseases), laboratory medicine (diseases of the thyroid gland, borreliosis, nephrology), internal medicine (rheumatology, diabetes mellitus, nephrology, hepatology), opthomology, physiology, and dentistry. A list of the selected applications are included in the Appendix. In a 1996 survey conducted by Kraut and Mann, goals for the purpose of patient care were pursued by 90% of the promoted projects, 8% attempted to improve patient care indirectly by supplying knowledge, and only 2% of all systems pursued goals exclusively associated with medical information; 85% of the developed systems include a component for decision support based on patient data, 65% include a knowledge-based encyclopedia, and 31% include a tutoring component. In addition, 54% of the projects provide "added values" (e.g., statistics about performance or reports) to improve the cost/benefit ratio. 5. RESEARCH ISSUESIn assessing applications, it is pertinent to examine the following research issues that affect the success of expert systems in medicine.
5.1. SELECTING A MEDICAL DOMAINThe ingredient for a successful application in medicine appears to be a careful choice of the medical domain. In selecting a medical domain, one should take into account whether the system is indispensable to users, the complexity and uncertainty in domain, and the number of users. The domain must be narrow and relatively self-contained. The system should provide substantive assistance to the physician, and the task should be one that the physician either cannot do without it or is willing to let a system do. As clinical domains become more complex, a number of difficult problems have tended to degrade the quality of performance of medical expert systems. Significant clinical problems require large knowledge bases that contain complex interrelationships, including time and functional dependencies. A particularly important feature in selecting a medical domain for designing an expert system is the desire to increase the quality of decision-making on the part of patients, physicians, or both. In the general practice of medicine, the quality of decision-making is diminished by the inability of humans, and physicians in particular, to deal effectively with uncertainty. Among other consequences, this inability promotes the abuse of diagnostic procedures by encouraging physicians to search for causal relationships without regard to the cost of their search (Holtzman, 1989). The attractiveness of a medical domain for an expert system is also affected by the routine usage and the use of major testing procedures. Examples of such procedures might include biopsies, endoscopies, amniocenteses, exploratory laparotomies, or any other diagnostic task having considerable financial cost, toxicity, invasiveness, and/or threat to life. These procedures increase the worth of a decision-analytic-based approach for selecting medical strategies because value-of-information calculations can be used very effectively to better make the selection. Finally, the number of users is another important factor in the success of a medical expert system. Consider, for example, the case of heart and lung transplant patients. Although the stakes involved in deciding whether or not to undergo these treatments are very high, the affected patient population is probably too small to make it worthwhile to develop a full-scale expert system. In these cases, individual decision analyses are likely to be more economical. 5.2. KNOWLEDGE ACQUISITION AND REPRESENTATIONCentral to ensuring a system's adequate performance is matching the most appropriate knowledge acquisition and representation methods with the problem domain. Good statistical data may support an effective Bayesian approach or other statistical method, such as discriminant analysis in settings where diagnostic categories are small in number, nonoverlapping, and well defined. But the inability to use qualitative medical knowledge limits the effectiveness of the statistical approach in more difficult patient management or diagnostic environments. On the other hand, neural network or CBR systems may support decision-making in complex and uncertain domains in which observations are typically quantified. Chae et al. (1996) compared three knowledge representation methods (namely, neural network, discriminant analysis, and CBR) for the diagnosis of asthma. The neural network had the best overall prediction rate (92%) and the best prediction rate for asthma (96%). The discriminant analysis had the best prediction rate for non-asthma (80%), and the CBR had the lowest prediction rates in all categories.
|
![]() |
|
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. |