Passive construction of diagnostic decision models: An empirical evaluation

Bayesian networks have proven their value in solving complex diagnostic problems. The main bottleneck in applying Bayesian networks to diagnosis is model construction. In our earlier work [1], we proposed passive construction of diagnostic models based on observation of diagnosticians solving diagnostic cases. This idea has never been tested in practice. In this paper, we describe an experiment that tests an interactive prototype system called MARILYN on implementation of a system based on passive construction of diagnostic model, by inputting four hundred help desk cases collected at the University of Pittsburgh campus computing lab. We show that while the system’s diagnostic accuracy continues to increase with the number of cases, it reaches very reasonable levels after merely tens of cases