As the variety of diagnostic and therapeutic modalities for cardiovascular disorders has expanded, the task of planning and assessing patient management has become increasingly difficult. Much of the reason for this difficulty stems from the many interacting physiological mechanisms that together determine the effect of any intervention in a particular disease context. Anticipating the mechanisms that will have the most important effects is a skill that requires considerable experience. The goal of the Heart Failure Program is to provide a tool that the practitioner can use for assistance in reasoning about the diagnostic and therapeutic questions involved in managing a patient[1]. While the program addresses many facets of the patient management process, this paper focuses on the development of a method to predict the response of a patient in a known disease state to a particular therapy or action.
The Heart Failure Program is organized around a causal physiological model of the cardiovascular system. This model is the knowledge base specifying the relationships among parameters in general as well as being the structure for recording the specific states and relationships in the patient. Around this physiological model are program modules for interpreting patient input, for reasoning about the possible diagnoses, and for reasoning about therapy. The problem of reasoning about therapy includes suggesting possible therapies for the patient state and explaining to the user the possible effects, both good and bad, that the therapies might have. Since the diagnosis is unlikely to completely reveal the state of the patient and our present understanding of physiology still allows considerable interpatient variation in responses, the predictions of therapy effects should include as much of the range of possibilities as is consistent with the known state of the patient. We are not yet at the stage of predicting the range of possible effects, but the thrust of the methodological development is toward reasoning about changes even when states and relationships are not completely determined and toward explaining predictions in terms of the mechanisms with the strongest influence on the results.