Using Artificial Intelligence to Develop Virtual Control Groups and Decision Support Tools for Toxicologists and Pathologists

抄録

Artificial intelligence (AI) based machine learning can facilitate data evaluation and 3R (refinement, reduction, replacement) initiatives for general toxicology studies supporting safety assessment. De-identified training and testing data sets were constructed from rat and non-human primate studies to develop AI based classifiers that could support 3 different intended uses: 1. Decision support for the identification of common pathology findings in 6 organs of the rat; 2. Decision support for the identification and etiology of a common hematology finding in the rat; and 3. Development of virtual control groups for the replacement or reduction of non-human primate control animals in general toxicology studies. Proof of concept classifiers were developed using both deep learning and traditional machine learning techniques. The classifiers were tested with studies not used for the training set and the output was evaluated statistically using standard methods as well as by experts (pathologists and toxicologists) for the intended use. The classifers showed good potential for a positive impact on data evaluation and outcomes associated with the 3Rs in general toxicology studies.