We review highlights of our research on Pathfinder, a decision-theoretic expert system for hematopathology diagnosis. We have developed techniques for efficiently acquiring, representing, and reasoning with uncertain biomedical knowledge. Specifically, we have developed a methodology for coping with complex dependencies among findings and disease in pathology. The methodology includes an extension of the belief-network representation called similarity networks. Using this methodology, we have constructed a large probabilistic knowledge base for the domain of lymph-node pathology. We have also developed techniques for improving the clarity of explanations through the use of human-oriented abstractions. Finally, we have conducted a formal evaluation of Pathfinder’s diagnostic accuracy. ∗Core research on the Pathfinder project has been supported by the National Library of Medicine under Grant RO1LM04529. Components of the research have been supported by the by the National Aeronautics and Space Administration under Grant NCC-220-51, by the National Science Foundation under Grant IRI8703710, and by the Josiah Macy, Jr. Foundation. Computational support has been provided by the SUMEX-AIM Resource under NIH Grant RR-00785.
Orthopaedic surgeons frequently refer patients to imaging centers for MR and CT studies in order to make a final decision about surgery or treatment. Typically, the radiologist studies the images of each case and dictates a full text report describing the findings. The text report is then transcribed, printed and returned to the radiologist for a final review and signature. At. the Laurie Imaging Center in New Brunswick, New Jersey, the full text report contains three parts: a description of the MRI study, a narrative description of the findings and an impression. The latter contains the essence of the interpretation, and constitutes the most relevant part of the report that affects the decisiOn to be made by the surgeon. There are several drawbacks to conventional text reports: writing them is cumbersome, thev contain ambiguities due to their textual nature, and they do not facilitate the development of well-defined outcome and retrospective studies. These lead to inefficiencies in communicating the radiological findings.
The broad objective was to develop an information system which integrates various sources of clinical data and facilities outcome assessment for patients evaluated in a lumbar spine service. During a patient encounter, the physician formulates a hypothesis regarding appropriate forms of treatment and he or she may then use this system to explore previous treatment outcomes for similar cases. The availability of a clinical tool that presents information in an outcome-oriented format may be highly relevant to the delivery of cost-efficient, high-quality health care and also create a formal mechanism for detecting practice variability.
The Regenstrief Medical Record System (RMRS) is used by three hospitals at the Indiana University Medical Center campus, and thirty off-campus clinic sites. It also operates at Wilford Hall Medical Center and Brooks Army base in San Antonio, Thomason Hospital in El Paso, Texas, and other sites. In Indianapolis, it captures data for about 60,000 hospitalizations and 600,000 outpatient encounters per year. Two thousand care providers (nurses, physicians, medical students) access RMRS data each month at Indiana University, Riley, and Wishard hospitals.
