T. Schiøler, W. Grimson, P. Sharpe, M. Egmont-Petersen, G. Momsen,
R. O'Moore, P. McNair. "Automatic decision support based on voting by independent
decision support systems," Proceedings of the CCL congress, Dublin,
pp. 58-66, 1992.
Abstract:
This paper compares the performance of a number of methods for automatic diagnosis of thyroid disease based on biochemical patient data. The methods studied were neural networks, probabilistic causal networks, induction algorithms and a "naive" manually constructed rule based system. All methods have been trained on the same database of 202 thyroid patients and normal controls. Since different methods diagnose a given test patient differently, a novel voting strategy is proposed which allows the diagnosis produced by each method to produce a single, majority diagnosis. Preliminary results on an independent test database of 174 patients shows that the voting strategy outperforms the individual methods on the basis of both correctness and coverage. This higher performance of the voting strategy is due to differences on which test cases the individual methods are able to perform and perform correct on.
Reprints, please contact me: michael@cs.uu.nl