A decision-theoretic network approach to treatment management and prognosis

Abstract

Medical patient management is a complicated process, usually involving a large amount of, possibly uncertain, information. Clinicians may, therefore, require some form of decision support to deal with complicated situations; assistance in exploring various clinical questions, e.g., concerning prognosis and optimal treatment, may be valuable in this respect. Decision-theoretic expert systems provide a suitable framework for such assistance due to the flexibility of the underlying formalisms, with inherent potentials of knowledge reuse. In this paper, the development of a decision-theoretic model of non-Hodgkin lymphoma of the stomach is described, and examined for its clinical usefulness. Central to the model is a probabilistic network that offers an explicit representation of the uncertainties underlying the decision-making process.

Introduction

Medical patient management is a complicated process, usually involving a large amount of, possibly uncertain, information. It is, therefore, not surprising that many computer-based systems, including expert systems, have been developed during the past two decades to offer assistance in diagnosis, treatment management and assessment of prognosis. The overall goal of the development of these systems was improvement of patient care. However, despite considerable research efforts, few of these systems have actually reached the stage of clinical use. This lack of success in the medical field has been attributed to a number of different factors, both technical and social [1]. This paper focuses on a factor that belongs to the technical realm of expert systems: potentials of reuse of medical knowledge.

Recent model-based approaches to the development of medical expert systems offer substantial advantages in terms of flexibility and potentials of knowledge reuse, mainly due to the declarative nature of the knowledge embodied in such systems. Frequently employed formalisms to represent knowledge in model-based systems include logic and probability theory, among others. In medicine, using probability theory, possibly augmented with decision theory, is attractive, because it allows for explicitly representing the uncertainties and preferences that go with much of the knowledge used in medical decision-making. In particular, adopting probabilistic networks as a framework for building medical expert systems is appealing, because in addition to the advantages linked to using probability theory, qualitative relationships among variables can be represented, in the form of a directed acyclic graph. This feature not only simplifies probability assessment, but also promotes easy understanding of the resulting system. We shall refer to a system that incorporates a probabilistic network, possibly augmented with decision theory, as a decision-theoretic expert system. A knowledge base of such a system will be called a decision-theoretic network.

There are presently no comprehensive knowledge-engineering methodologies available that can be used in building a decision-theoretic network. In this paper, a further elaboration of a preliminary methodology for building decision-theoretic expert systems for treatment management, as described in [2], is discussed. It is also investigated how decision-theoretic expert systems can be used to deal with a number of questions clinicians are likely to find important: optimal treatment selection, prognostic assessment and generating profiles for specific treatment outcomes. Since all of these questions can be answered, in principle, using the same knowledge base of a decision-theoretic expert system, a practical example of knowledge reuse is arrived at.

As an example, medical treatment of patients with non-Hodgkin lymphoma of the stomach (gastric NHL) is discussed. Primary gastric NHL is a relatively rare malignant disorder, accounting for about 5% of gastric tumours. Until recently, the aetiology of gastric NHL was unknown; it is now generally believed that the main factor in the pathogenesis of this disease is a chronic infection with the bacterium Helicobacter pylori 3, 4. Therapy selection for gastric NHL is a complicated process, because only part of the patient findings necessary for therapy selection may be known at a particular stage of the disease, and knowledge of adverse reactions to particular treatments in patient groups may influence treatment selection significantly.

The structure of this paper is as follows. First, we discuss the process of building a decision-theoretic network, which is illustrated by the construction of a decision-theoretic network for gastric NHL. Next, various possible clinical applications of the resulting decision-theoretic expert system are considered. Finally, decision-theoretic networks are compared with conventional prognostic models in medicine.