Impact of Quality of Bayesian Network Parameters on Accuracy of Medical Diagnostic Systems

While most knowledge engineers believe that the quality of results obtained by means of Bayesian networks is not too sensitive to imprecision in probabilities, this remains a conjecture with only modest empirical support. We summarize the results of several previously presented experiments involving Hepar II model, in which we manipulated the quality of the model’s numerical parameters and checked the impact of these manipulations on the model’s accuracy. The chief contribution of this paper are results of replicating our experiments on several medical diagnostic models derived from data sets available at the Irvine Machine Learning Repository. We show that the results of our experiments are qualitatively identical to those obtained earlier with Hepar II.