On the Sensitivity of Probabilistic Networks to Reliability Characteristics

Abstract

This chapter examines a sensitivity analysis of a real-life probabilistic network undertaken to determine the effects of inaccuracies in the reliability characteristics of diagnostic tests. A sensitivity analysis of the esophageal cancer network was undertaken. In this analysis, the parameter probabilities of all test variables discerned were varied, and the effects of their variation on the most likely stage computed from the network were analyzed. Because the patterns of sensitivity exhibited by a network typically vary with evidence, the medical records from 185 patients diagnosed with cancer of the esophagus were used. The analysis revealed that for esophageal cancer, the prior probability of hematogenous metastasis being present was relatively small. Upon diagnostic reasoning, moreover, it would not increase unless there is some strong evidence of metastases in a patient's liver or lungs. If varying a parameter probability induces a change in the most likely value of the main diagnostic variable of a network, then inaccuracies in the parameter are likely to affect the network's diagnosis. The extent to which such inaccuracies can be influential is expressed by the admissible deviation for the parameter probability under study.