Programming Research Group Technical Monograph PRG-83

Bryan S. Todd

May 1990, 15 pages, ISBN 0-902928-62-7

The statistical analysis of collections of previous case records has proved a useful way of giving diagnostic assistance to the clinician. In certain applications, "simulation models" of disease processes provide a way of supplementing the available numerical data with the causal relationships that are known to exist. However, the diagnosis of new patients by reference to such simulation models tends to be computationally hard. In these circumstances a possible solution is to use the model to generate randomly a database of hypothetical cases which is sufficiently large to enable a more effective form of statistical classification than was previously possible. In this paper, several classifiers are considered for this purpose. A method is described for comparing the diagnostic accuracy of the classifiers in a way which is independent of the medical correctness of the simulation model itself. The method is illustrated by an example.