Quantitative Risk Assessment of the Pulmonary Toxicity of Nano- particles by Machine-Learning-Enabled Meta-Analysis

Accurately anticipating the toxic risks and specific factors contributing to the toxic risks of nanomaterials is a necessary step for the safe and effective proliferation, utilization, and regulation of these unique materials. This thesis addresses this problem through meta-analysis on existing nanomaterial pulmonary toxicity experiments as enabled by the use of machine learning algorithms including regression trees and random forests models at a time when the completeness of the data do not support traditional meta-analysis techniques like multiple linear regression. This thesis presents the results of analysis using these models to identify the most important nanomaterial characteristics contributing to toxicity as well as the magnitude of changes in toxicity expected from changes in those characteristics.

This thesis presents predictive models for the pulmonary toxicity of carbon nanotubes and titanium dioxide nanoparticles showing the degree to which changes in experimental design, nanomaterial dimensions, impurities, and aggregation might explain differences in observed toxicity.

Secondly, this thesis presents the predictions of random forest models revealing interactions between 2 or 3 nanomaterial characteristics and exposure attributes in a manner such that a material designer might minimize risk while continuing to meet functional objectives.