We seek support to (1) investigate fundamental issues in applications of artificial intelligence to medicine within the context of lymph node pathology and (2) construct a useful diagnostic system for the hematopathology field. The complexity of hematopathology poses a rich set of challenges for artificial intelligence in medicine. Our research addresses fundamental problems of knowledge representation, reasoning strategies, user modeling, explanation, and user acceptance. Our goal is to build a useful decision system for pathologists. Such a working system would be extraordinarily useful for the lymph node pathology field. Many studies have documented that there are major problems in the diagnosis and classification of malignant lymphomas and they are often confused with benign diseases. This poses a major problem for the oncologist as accurate diagnosis is crucial for appropriate treatment. We have built a prototype experty system called PATHFINDER which makes diagnoses of lymph node pathology through a consideration of 50 malignant and 30 benign diseases and approximately 500 histopathologic findings. The program uses the method of sequential diagnosis; it can generate differential diagnoses based on a small number of findings and can suggest other findings useful for narrowing the differential diagnosis. Consensus meetings have already resulted in agreement on features and diseases to be used by the program. Our research plans are to further develop PATHFINDER so that it can become not only a regularly used diagnostic tool for users having different levels of training and experience, but also can serve as a powerful teaching system. The proposed expert system will provide a unique integrated knowledge base for the benign diseases of lymph nodes. AIDS, and lymphomas giving the practicing pathologist immediate access to accurate morphologic classification and diagnostic criteria, while at the same time providing guidance as to whether additional tests (immunologic, cytogenetic, cell-kinetic, and immunogenetic) are likely to be informative. It is anticipated that a working diagnostic system could improve the reproducibility and reliability of lymphoma diagnosis, facilitate the standardization of diagnoses in clinical trials, and help ensure appropriate therapy at the community hospital level. We plan to substantially increase the quantity and quality of knowledge through (1) the integration of independent knowledge bases from the four experts mentioned above, (2) the integration of laboratory, clinical, immunologic, cell kinetic, cytogenetic, and immunogenetic information, and (3) the restructuring of the knowledge base to include dependencies among features. Furthermore, we will (4) facilitate the identification of morphologic features and standardize definitions of features and diseases by incorporating detailed textual and pictorial descriptions (using videodisk technology), (5) develop mapping schemes that will facilitate the understanding of the many classification systems for non-Hodgkin's lymphomas, and (6) implement methods for the detection of multiple discordant histologies, progression of disease over time, and atypical proliferation. Also, we will implement and incorporate into the program (1) scoring methods when features are interdependent, (2) heuristics for high-level control of reasoning strategies, (3) heuristics which customize the bahavior of the program to the expertise of the user, (4) formal decision analysis techniques, and (5) methods for system explanation. Throughout the three year period, we will also monitor our progress through iterative testing of reasoning strategies and frequent validations of the program's knowledge.
