The goal of this research is to design and build a computer program which can understand a realistic case description, construct a detailed model of the internal state of the patient and its evolution, and answer questions about the pathophysiology of the particular case. This research addresses a need for medical reasoning programs that can demonstrate a deep understanding of the state of the patient, rather than simply producing the name of a disease. Our current research has produced a representation for causal relationships in physiological mechanisms that includes a qualitative language for describing the structure of a mechanism and a method for envisioning the possible behaviors of that mechanism. We propose to use this knowledge representation as the basis for a program to reason about a limited, but natural and complex medical domain: disorders of sodium and water balance. The domain of sodium and water balance is well-understood scientifically, but poses complex clinical problems that will be a realistic test for the representation. The program will assimilate information about a patient in the format of a natural case description, and will determine the diagnosis and formulate a detailed description of the patient's internal state. We will evaluate the program's behavior according to its ability to answer the types of questions that might be asked of a resident during clinical training. Our research method rests heavily on the detailed observation of human experts, both to capture the knowledge used and to determine the representation for that knowledge.
