Fuzzy complex number aided evaluation of predictive toxicology models

There is a growing interest in applying computational intelligence in the predictive toxicology (PT) domain, where a large number of predictive models are becoming available. Evaluation of such models is therefore considered to be a crucial part of their development and potential use, especially for regulatory purposes. The current evaluation approaches mainly focus on statistical measures of model performance, and few of them have taken data quality into consideration. However, it has been well recognised that datasets and models should not be considered in isolation. This paper proposes a new confidence index for evaluating PT models. A fuzzy complex number (FCN) framework is expanded in an effort to represent and evaluate dataset and regression-based model quality in a two-dimensional manner, thereby ensuring the linguistic evaluation is transparent and explainable. The utility and applicability of this research is illustrated by an experiment which evaluates 17 regression-based PT models. The experimental results have been compared and analysed against existing methods, and show that the FCN-based approach provides a consistent and interpretable means of model assessment. The proposed indexing mechanism can be used, together with customised statistical measures, in assisting PT model selection. This approach also helps to capture the relationships between datasets and models, and contributes to the development of data and model governance in PT.