Edward H. Shortliffe

A model of inexact reasoning in medicine

Abstract Medical science often suffers from having so few data and so much imperfect knowledge that a rigorous probabilistic analysis, the ideal standard by which to judge the rationality of a physicians decision, is seldom possible. Physicians nevertheless seem to have developed an ill-defined mechanism for reaching decisions despite a lack of formal knowledge regarding the interrelationships of all the variables that they are considering. This report proposes a quantification scheme which attempts to model the inexact reasoning processes of medical experts. The numerical conventions provide what is essentially an approximation to conditional probability, but offer advantages over Bayesian analysis when they are utilized in a rule-based computer diagnostic system. One such system, a clinical consultation program named mycin , is described in the context of the proposed model of inexact reasoning.

Production rules as a representation for a knowledge-based consultation program

The MYCIN system has begun to exhibit a high level of performance as a consultant on the difficult task of selecting antibiotic therapy for bacteremia. This report discusses issues of representation and design for the system. We describe the basic task and document the constraints involved in the use of a program as a consultant. The control structure and knowledge representation of the system are examined in this light, and special attention is given to the impact of production rules as a representation. The extent of the domain independence of the approach is also examined.

Clinical decision-support systems

If you ask people what the phrase “computers in medicine” means, they often describe a computer program that helps physicians to make diagnoses. Although computers play numerous important medical roles, from the earliest days of computing people have recognized that computers might support physicians by helping these people to sift through the vast collection of possible diseases and symptoms. This idea has been echoed in futuristic works of science fiction. In Star Trek, for example, medical workers routinely point devices at injured crew members to determine instantly what is the problem and how serious is the damage. The prevalence of such expectations, coupled with a general societal concern about the influence of computers on interpersonal relationships and on job security, has naturally raised questions among health workers. Just what can computers do today to support clinical decision-making? How soon will diagnostic tools be generally available? How good will they be? What will their effects be on the practice of medicine, on medical education, and on relationships among colleagues or between physicians and patients?

A method for managing evidential reasoning in a hierarchical hypothesis space

Although informal models of evidential reasoning have been successfully applied in automated reasoning systems, it is generally difficult to define the range of their applicability. In addition, they have not provided a basis for consistent management of evidence bearing on hypotheses that are related hierarchically. The Dempster–Shafer (D-S) theory of evidence is appealing because it does suggest a coherent approach for dealing with such relationships. However, the theory’s complexity and potential for computational inefficiency have tended to discourage its use in reasoning systems. In this paper we describe the central elements of the D-S theory, basing our exposition on simple examples drawn from the field of medicine. We then demonstrate the relevance of the D-S theory to a familiar expert-system domain, namely the bacterial-organism identification problem that lies at the heart of the mycin system. Finally, we present a new adaptation of the D-S approach that achieves computational efficiency while permitting the management of evidential reasoning within an abstraction hierarchy

The GuideLine Interchange Format: A Model for Representing Guidelines

OBJECTIVE To allow exchange of clinical practice guidelines among institutions and computer-based applications. DESIGN The GuideLine Interchange Format (GLIF) specification consists of GLIF model and the GLIF syntax. The GLIF model is an object-oriented representation that consists of a set of classes for guideline entities, attributes for those classes, and data types for the attribute values. The GLIF syntax specifies the format of the test file that contains the encoding. METHODS Researchers from the InterMed Collaboratory at Columbia University, Harvard University (Brigham and Womens Hospital and Massachusetts General Hospital), and Stanford University analyzed four existing guideline systems to derive a set of requirements for guideline representation. The GLIF specification is a consensus representation developed through a brainstorming process. Four clinical guidelines were encoded in GLIF to assess its expressivity and to study the variability that occurs when two people from different sites encode the same guideline. RESULTS The encoders reported that GLIF was adequately expressive. A comparison of the encodings revealed substantial variability. CONCLUSION GLIF was sufficient to model the guidelines for the four conditions that were examined. GLIF needs improvement in standard representation of medical concepts, criterion logic, temporal information, and uncertainty.

Probabilistic similarity networks

I address practical issues concerning the construction of normative expert systems--expert systems that encode knowledge within a decision-theoretic framework. In particular, I examine the similarity network and partition, two extensions to the influence diagram. A similarity network is a tool for building an influence diagram, whereas a partition is a tool for assessing the probabilities associated with an influence diagram. Both representations encode asymmetric forms of conditional independence that are not represented conveniently in an ordinary influence diagram. Similarity networks and partitions exploit these forms of conditional independence to facilitate the construction and assessment of influence diagrams for problems of diagnosis. The representations aided considerably the construction of Pathfinder, a large normative expert system for the diagnosis of lymph-node diseases (the domain contains approximately 60 diseases and 110 disease findings). In an early version of the system, I encoded the knowledge of the expert using an erroneous assumption that all disease findings were conditionally independent, given each disease. When the expert and I attempted to build an influence diagram for the domain to capture the dependencies among the disease findings, we failed. Using a similarity network, however, we were able to construct the influence diagram for the entire domain in approximately 40 hours. Furthermore, using the partition representation, the expert was able to decrease the time required to assess a probability--on average--by almost one order of magnitude. Most important, through a comparison procedure based in decision theory, I found that the improvements in diagnostic accuracy afforded by the more sophisticated model of the domain were well worth the additional effort that we had invested to build the revised version of the system. In this work, I examine in detail the theoretical properties of similarity networks and partitions, and discuss the application of these representations to the construction of Pathfinder. This research suggests strongly that, by identifying specific forms of conditional independence, and by developing representations that exploit these forms of independence for knowledge acquisition, knowledge engineers can construct normative expert systems for domains of larger scope and greater complexity than the domains previously through to be amenable to the decision-theoretic approach.

Knowledge engineering for medical decision making: A review of computer-based clinical decision aids

Computer-based models of medical decision making account for a large portion of clinical computing efforts. This article reviews representative examples from each of several major medical computing paradigms. These include 1) clinical algorithms, 2) clinical databanks that include analytic functions, 3) mathematical models of physical processes, 4) pattern recognition, 5) Bayesian statistics, 6) decision analysis, and 7) symbolic reasoning or artificial intelligence. Because the techniques used in the various systems cannot be examined exhaustively, the case studies in each category are used as a basis for studying general strengths and limitations. It is noted that no one method is best for all applications. However, emphasis is given to the limitations of early work that have made artificial intelligence techniques and knowledge engineering research particularly attractive. We stress that considerable basic research in medical computing remains to be done and that powerful new approaches may lie in the melding of two or more established techniques.

GLIF3: a representation format for sharable computer-interpretable clinical practice guidelines

The Guideline Interchange Format (GLIF) is a model for representation of sharable computer-interpretable guidelines. The current version of GLIF (GLIF3) is a substantial update and enhancement of the model since the previous version (GLIF2). GLIF3 enables encoding of a guideline at three levels: a conceptual flowchart, a computable specification that can be verified for logical consistency and completeness, and an implementable specification that is intended to be incorporated into particular institutional information systems. The representation has been tested on a wide variety of guidelines that are typical of the range of guidelines in clinical use. It builds upon GLIF2 by adding several constructs that enable interpretation of encoded guidelines in computer-based decision-support systems. GLIF3 leverages standards being developed in Health Level 7 in order to allow integration of guidelines with clinical information systems. The GLIF3 specification consists of an extensible object-oriented model and a structured syntax based on the resource description framework (RDF). Empirical validation of the ability to generate appropriate recommendations using GLIF3 has been tested by executing encoded guidelines against actual patient data. GLIF3 is accordingly ready for broader experimentation and prototype use by organizations that wish to evaluate its ability to capture the logic of clinical guidelines, to implement them in clinical systems, and thereby to provide integrated decision support to assist clinicians.

Computer-based consultations in clinical therapeutics: Explanation and rule acquisition capabilities of the MYCIN system☆

Abstract This report describes progress in the development of an interactive computer program, termed MYCIN, that uses the clinical decision criteria of experts to advise physicans who request advice regarding selection of appropriate antimicrobial therapy for hospital patients with bacterial infections. Since patients with infectious diseases often require therapy before complete information about the organism becomes available, infectious disease experts have identified clinical and historical criteria that aid in the early selection of antimicrobial therapy. MYCIN gives advice in this area by means of three subprograms: (1) A Consultation System that uses information provided by the physician, together with its own knowledge base, to choose an appropriate drug or combination of drugs; (2) An Explanation System that understands simple English questions and answers them in order to justify its decisions or instruct the user; and (3) A Rule Acquisition System that acquires decision criteria during interactions with an expert and codes them for use during future consultation sessions. A variety of human engineering capabilities have been included to heighten the programs acceptability to the physicians who will use it. Early experience indicates that a sample knowledge base of 200 decision criteria can be used by MYCIN to give appropriate advice for many patients with bacteremia. The system will be made available for evaluation in the clinical setting after its reliability has been shown to approach that of infectious disease experts.

PUFF: an expert system for interpretation of pulmonary function data

The application of artificial intelligence techniques to real-world problems has produced promising research results, but seldom has a system become a useful tool in its domain of expertise. Notable exceptions are the DENDRAL (1) and MOLGEN (2) systems. This paper describes PUFF, a program that interprets lung function test data and has become a working tool in the pulmonary physiology lab of a large hospital. Elements of the problem that paved the way for its success are examined, as are significant limitations of the solution that warrant further study.