MEDICAL DIAGNOSIS USING A PROBABILISTIC CAUSAL NETWORK
Applied Artificial Intelligence
This paper relates our experience in developing a mechanism for reasoning about the differential diagnosis of cases involving the symptoms of heart failure using a causal model of the cardiovascular hemodynamics with probabilities relating cause to effect. Since the problem requires the determination of causal mechanism as well as primary cause, the model has many intermediate nodes as well as causal circularities requiring a heuristic approach to evaluating probabilities. The method we have
... e method we have developed builds hypotheses incrementally by adding the highest probability path to each finding to the hypothesis. With a number of enhancements and computational tactics, this method has proven effective for generating good hypotheses for typical cases in less than a minute. Acknowledgements The research reported here was done with my medical collaborators a t New England Medical Center, Shapur Naimi, M. G. Criscitiello, Stephen Pauker, Greg Larsen, and in earlier years of the project, Robert Jayes and Steven Kurzrok. They are responsible for the medical content of the program, although any mistakes in this paper are my own. The work reported here has been supported by National Institutes of Health grants R01 HL 33041 from the National Heart, Lung, and Blood Institute and R01 LM 04493 from the National Library of Medicine. Over the past few years we have been developing a system to assist the physician in the diagnosis and management of patients with diseases that may cause or resemble heart failure. The program will be used by interns, residents, house staff, and other physicians who are managing patients with complex disorders over a period of days in a setting such as an intensive care unit. It will assist the physician in reasoning about the patient for both diagnosis and management. The physician can enter information about the history, physical examination, and laboratory tests (i.e., the findings) and the program will provide a differential diagnosis list with a graphical explanation of how each set of causes could produce the findings, suggestions about what other information would help to differentiate among the possibilities, suggestions about therapies that could correct the causal paths leading to undesirable states, and provide predictions about the overall effects of various therapies given the patient's pathophysiological state. The focus of this paper is the problem of providing useful information about the diagnosis. The cardiovascular domain, as well as many others in medicine, is full of uncertain causal mechanisms. This has lead to a representation of the domain knowledge as a network of causal probabilistic links between physiologic parameter states. However, this network contains multiple paths between nodes and forward cycles, necessitating the development of heuristic methods for evaluating the state of the network when findings are known. These methods are used to produce multiple hypotheses representing possible explanations for the findings, which are put together as a differential diagnosis. We have found that these methods are effective for interactive use in a model that contains about 150 internal nodes and about 300 possible findings. The following sections will discuss the special problems of the medical domain, the nature of a differential diagnosis, the kinds of causal relationships and the way they are modeled as probabilities in the program, the approach to producing diagnostic hypotheses, and an example of its use. The emphasis throughout is on the lessons we have learned about the nature of the problems and the practical concerns in building tools to meet the needs of the users.