Optimal Diagnostic Testing Strategies via Influence Diagrams
*Jarrod E. Dalton, Cleveland Clinc
Keywords: Bayesian networks; influence diagrams; diagnostic testing
We present a novel application of Bayesian networks (BNs) and influence diagrams (IDs) for medical diagnosis with the potential to reduce excess utilization of diagnostic tests. BNs are statistical models for causal systems, characterized by directed acyclic graphs; each node represents a random variable and each directed link represents pairwise causality among random variables. A variety of inference algorithms allow for dynamic updating of beliefs about disease in a given patient as tests are performed. IDs are BNs with additional nodes representing decisions and utility measures. Model construction involves baseline characteristics as parents of the disease node, tests as children of the disease node, a treatment decision node as a child of the disease node, and a utility node defined as a function of the treatment decision node and disease status node. After populating the ID, probability-of-disease thresholds for stopping testing are estimated based on maximizing expected utility, and for any given combination of observed test nodes, an ordering of future tests that maximizes expected utility is identified through Markov Chain Monte Carlo approximation.