The ALARM Monitor and the Bone-Marrow Transplant Therapy Advisor: A Demonstration of Two Probabilistic Expert Systems in KNET

摘要

ALARM (A Logical Alarm Reduction Mechanism) is a diagnostic application used to explore probabilistic reasoning techniques in belief networks. ALARM implements an alarm message system for patient monitoring; it calculates probabilities for a differential diagnosis based on available evidence [1]. The medical knowledge is encoded in a graphical structure connecting 8 diagnoses, 16 findings and 13 intermediate variables.The goal of the ALARM monitoring system is to provide specific text messages advising the user of possible problems. This is a diagnostic task, and we have chosen to represent the relevant knowledge in the language of a belief network. This graphical representation [6] facilitates the integration of qualitative and quantitative knowledge, the assessment of multiple faults, as required by our domain, and nonmonotonic and bidirectional reasoning.We have also created a belief network, the Bone-Marrow Transplant Therapy Advisor, that represents prognostic factors and their effects on possible outcomes of a bone-marrow transplant. For pediatric patients in the advanced stages of acute lymphoblastic leukemia (ALL), bone-marrow transplantation is generally considered the most promising therapy. For the patient and parents, the decision to proceed with transplantation is often difficult. Morbidity after transplantation is usually severe, and a significant percentage of those who receive a bone marrow transplantation die within a year of transplantation [7]. Many factors, however, offer significant insight into the expected outcome of marrow transplantation. A few examples of such prognostic factors include the white blood count at diagnosis, the age at diagnosis, the number of recurrence episodes before transplantation, and the quality of the match with the marrow donor. Some of those factors indicate the progress of the disease, whereas others define sensitivity to the chemotherapeutic conditioning regimee or the likelihood of Graft-versus-Host Disease (GvHD).Within the discipline of medical informatics, many researchers have studied methodologies for encoding the knowledge of expert clinicians as computational artifacts. KNET, the support software for ALARM and the bone-marrow transplant advisor, is a general-purpose environment for constructing probabilistic, knowledge-intensive systems based on belief networks and decision networks [2]. KNET differs from other tools for expert-system construction in that it combines a direct-manipulation visual interface with a normative, probabilistic scheme for the management of uncertain information and inference. The KNET architecture defines a complete separation between the hypermedia user interface on the one hand, and the representation and management of expert opinion on the other.In our laboratory, we and others have used KNET to build not only the ALARM and bone-marrow transplant systems, but also consultation programs for lymph-node pathology and clinical epidemiology [2,4]. KNET imposes few restrictions on the interface design. Indeed, we have rapidly prototyped several direct-manipulation interfaces that use graphics, buttons, menus, text, and icons to organize the display of static and inferred knowledge. The underlying normative representation of knowledge remains constant.We present ALARM and the transplant therapy advisor as part of a suite of probabilistic, knowledge-intensive medical expert systems. Such systems• Manage large quantities of extensively cross-referenced information• Emphasize clarity in acquiring, storing, and displaying expert knowledge• Incorporate tools for building hypertext user interfaces• Impose a limited number of constraints on the knowledge engineer's design choices• Share an axiomatic grounding for diagnosis and decision-making in probability theory and utility theory• Make normatively correct decisions and diagnoses in the face of uncertain, incomplete, and contradictory information• Draw inferences from knowledge bases large enough to model significant, real-world medical domains, and do so in polynomial time on low-cost hardwareIn this demonstration, we show how ALARM and the therapy advisor synthesize physiologic measurements and prognostic indicators into a diagnostic conclusion according to a belief-network model of the domains. We demonstrate KNET's hypertext interface and the transparent integration of probabilistic reasoning into a diagnostic application. KNET runs on any Macintosh II personal computer with at least 4 megabytes of random-access memory. The authors will provide all the necessary software on a SCSI hard disk. KNET fully supports color and monochrome monitors of any size, and requires no special hardware. We prefer, but do not require, a large color monitor, which demonstrates the capabilities of KNET to greatest advantage.