While the goal of diagnosis is to determine the pathophysiology in sufficient detail to suggest a rational approach to therapy, this goal usually can not be achieved from the initial history, presenting complaints, and physical examination. Tests or reexamination may be needed to determine the appropriate diagnosis. The tool for directing the diagnostic process is the differential diagnosis. The differential diagnosis identifies possible explanations for the known findings, identifies states that have been ruled out, and provides the data needed to determine what information can be used to refine the diagnosis.
Over the course of this research our view of differential diagnoses has evolved. Initially, we used the model and findings to identify parameter states that were well supported. For example, pulmonary findings would be used to decide whether pulmonary congestion was true or false in the patient. Then these states (e.g., pulmonary congestion) could be used as evidence for other parameter states. In this way the diagnosis could be built state by state. At any point in the process, the differential diagnosis was the union of the possible values of the states not yet known with the states known to be true. This is essentially the approach that was taken by CASNET[Weiss78] to reason from a causal network. There are two problems with this approach. First, it assumes there will always be enough evidence local to a parameter state (easily computable causes or consequences) to determine its truth, either immediately or given more test results. In fact, what often happens is that there are parameters with states whose probabilities are not overwhelmingly true or false, but any decision about one parameter state will change the probability of the others with possible far reaching consequences. For example, a minor decision of whether the right ventricle has decreased compliance versus fixed capacity in the model could determine whether the best diagnosis is restrictive cardiomyopathy or constrictive pericarditis - a minor distinction on which to make the diagnosis. Second, the view of the differential as all unknown states does not give a good picture of the likelihood of different possibilities, especially the significant combinations of states that would have consequences in patient management.
Because of the restrictions of the parameter by parameter view of the diagnosis problem, we started considering complete hypotheses - hypotheses that include causal chains from the primary causes (causes that do not require a further cause) to the findings and that cover all of the patient findings. With these it is possible to get a complete picture of what might be happening in the patient. This approach also allows the user to evaluate the reasoning of the computer. Since the hypotheses are complete, there are no hidden assumptions or implications. The user can decide whether the scenario proposed by the computer is reasonable and if not, why not.
Given that a hypothesis is complete from primary cause to findings, we are still left with the problem of how to form a differential. The set of all possible hypotheses is much too large to list and the hypothesis with the highest probability gives no indication of the alternatives. One could generate alternative hypotheses by taking the best hypothesis and proposing alternate paths for single parameter states or short chains of states. However, many of the alternatives that might be generated would differ from the starting hypothesis in inconsequential ways, while major alternatives would be overlooked. We have addressed the issue of what constitutes important alternatives by designating a subset of the parameter states (nodes) in the model as diagnostic nodes. Hypotheses are significantly different if they contain a different set of diagnostic nodes. This heuristic has proven effective in limiting the differential diagnosis to a reasonable number of alternatives that have significant differences.