Experimental analysis of large belief networks for medical diagnosis

We present an experimental analysis of two parameters that are important in knowledge engineering for large belief networks. We conducted the experiments on a network derived from the Internist-1 medical knowledge base. In this network, a generalization of the noisy-OR gate is used to model causal independence for the multivalued variables, and leak probabilities are used to represent the nonspecified causes of intermediate states and findings. We study two network parameters, (1) the parameter governing the assignment of probability values to the network, and (2) the parameter denoting whether the network nodes represent variables with two or more than two values. The experimental results demonstrate that the binary simplification computes diagnoses with similar accuracy to the full multivalued network. We discuss the implications of these parameters, as well other network parameters, for knowledge engineering for medical applications.