ATLA::Alternatives to Laboratory Animals

Volume 28, Number 5

Local irritation/corrosion testing strategies: extending a decision support system by applying self-learning classifiers.

ATLA 28, 651-663, September/October 2000

Stephan Zinke and Ingrid Gerner

Federal Institute for Health Protection of Consumers and Veterinary Medicine (BgVV), Thielallee 88-92, 14195 Berlin, Germany

SUMMARY

Procedures have been established and tested for the extension of a decision support system (DSS) for the prediction of the local irritation/corrosion potential of chemicals by using self-learning classifiers. The different approaches (decision tress, distances examinations in a multidimensional space, k-nearest neighbour method) have been implemented, tested and evaluated independently. A combination of all of the established extension approaches was also developed and tested. Self-learning classifiers are constructed "automatically" by a computer, i.e. they are not derived by a human expert, and thus they can be constructed with minimal effort. The classifiers presented here extend the existing DSS in a manner that increased significantly the predictive power of the extended system. However, automatically calculated results of self-learning classifiers are produced by a machine, and a machine is incapable of explaining the toxicological relevance of the results obtained. Thus, these results must be accepted, despite an inability to prove their reliability. Only the mathematical correctness of the method and the prediction rates for suitable test cases can lend some credibility to predictions produced by a computer calculating on a self-learning basis. This may not be adequate for regulatory hazard assessment purposes.

Keywords: decision support systems, self-learning classifiers, artifical intelligence methods, deicision trees, k-nearest neighbour method