Development of a hybrid Bayesian network model for predicting acute fish toxicity using multiple lines of evidence

Abstract

A Bayesian network was developed for predicting the acute toxicity intervals of chemical substances to fish, based on information on fish embryo toxicity (FET) in combination with other information. This model can support the use of FET data in a Weight-of-Evidence (WOE) approach for replacing the use of juvenile fish. The BN predicted correct toxicity intervals for 69%-80% of the tested substances. The model was most sensitive to components quantified by toxicity data, and least sensitive to components quantified by expert knowledge. The model is publicly available through a web interface. Further development of this model should include additional lines of evidence, refinement of the discretisation, and training with a larger dataset for weighting of the lines of evidence. A refined version of this model can be a useful tool for predicting acute fish toxicity, and a contribution to more quantitative WOE approaches for ecotoxicology and environmental assessment more generally.\n\nHighlightsO_LIA Bayesian network (BN) was developed to predict the toxicity of chemicals to fish\nC_LIO_LIThe BN uses fish embryo toxicity data in a quantitative weight-of-evidence approach\nC_LIO_LIThe BN integrates physical, chemical and toxicological properties of chemicals\nC_LIO_LICorrect toxicity intervals were predicted for 69-80% of test cases\nC_LIO_LIThe BN is publicly available for demonstration and testing through a web interface\nC_LI