Influence of Data Preprocessing and Kernel Selection on Probabilistic Neural Network Modeling of the Acute Toxicity of Chemicals to the Fathead Minnow and Vibrio fischeri Bacteria

Abstract

Abstract We investigated the connection between the data preprocessing strategy and kernel choice on the quality of the associated basic probabilistic neural network models for the acute toxicity of various chemicals to the fathead minnow and to Vibrio fischeri bacteria. The models employ exclusively structural parameters and physicochemical properties as inputs. Results show that the Gaussian kernel is preferable over the reciprocal kernel model. Data preprocessing based on the hyperbolic tangent and the sigmoid logistic transforms provides the best results at the level of the cross validation experiment. Improved models based on cross validation partial models and linear corrections were also investigated. The results show that the improved models with data preprocessing based on the hyperbolic tangent and the finite interval transforms are the best with practically identical quality of predictions.