Risk-Sensitive Diagnosis and the Role of Neural Networks

Diagnostic problem solving, whether it be fault-diagnosis in an engineering system or diagnosis of a disease in human beings, is a prime example of decision making in the face of uncertainty. Frequently, many different outcomes may correspond to an identical set of measured data or symptoms. The converse may also be true, that any given diagnosis may correspond to a number of distinct sets of diagnostic data. In addition, the data themselves may be imprecise adding to the overall uncertainty in the reasoning process, making it probablistic in nature. These factors can often be the cause of poor diagnostic accuracy and in part responsible for the difficulty in developing useful and usable diagnostic support systems. Furthermore, it would be unusual for diagnostic errors to be viewed as equally acceptable. For example, a large number of false alarms may be tolerable in the dignosis of heart attack when the decision to be made is simply admit to hospital or not. The level of acceptability changes though, when the decision to be made is whether or not to administer potentially life-threatening drugs. Evidently, the risk associated with an incorrect diagnosis is crucial to making a decision about treatment............