The role of Bayesian networks in the diagnosis of pulmonary embolism

Pulmonary embolism (PE) is a life-threatening condition and, despite advances in diagnostic technology, it remains an elusive diagnosis. The rich variety of possible clinical presentations make it particularly difficult to represent the diagnostic process as a 'decision tree'. However, Bayesian networks offer the opportunity of a compact representation of the domain underlying the decision process: once the network portrays the natural history of the disease, the utility of investigations can be quantitatively evaluated. We developed a network for the diagnosis of PE, including 72 variables to represent both the risk factors and the pathophysiological consequences of the disease. Its structure has been specified by discussing which causal relationships explain the manifestations of the disease. The quantitative measures of associations were retrieved from the medical literature, through a critical review of available studies and agreement on the assumptions made to cope with the lack of published information. Six examples are presented to illustrate the appropriateness of 'entropy reduction' as a measure of the utility of investigation, once the set of examinations to be evaluated is bounded on the grounds of their cost and the patient's current risk. The network, which has been given the acronym 'BayPAD' (Bayesian network for Pulmonary embolism Assisted Diagnosis), appears to be able to detect which observations make others irrelevant, so that decisions can be tailored to single cases.