A Computationally Efficient Approach to Predict Population Risk with Machine Learning

AbstractThe growing use of e-cigarettes and vaping devices in recent years is a concern for the health community. While the safety has not yet been fully characterized, these devices are linked to smoking cessation efforts, targeted marketing campaigns towards adolescents, and additives, such as fruit flavors, that promote use. Experimental data has been collected to investigate toxicity, lethality, and risk for cancer. However, the gaps in this type of data and the difficulty collecting large datasets leads to challenges with risk assessment calculations. Computational modeling to predict chemical and toxin distribution, deposition, and dosimetry has been successfully demonstrated; however, the computational requirements are prohibitive for large population studies. We hypothesize that replacing expensive computational models with a machine learning model will produce accurate risk assessment for a low computational cost and that this process can be generalized for other environmental health data.This project is a close collaboration between Kitware, Inc. and Applied Research Associates, Inc. (ARA). The Kitware team has extensive experience developing computational physiology models for use in simulation, storage, curation, and analysis of large dataset for medical and health related analysis, and machine learning techniques. We have developed an open source platform, Girder, for creating customized workflows related to large datasets and machine learning analysis. ARA has extensive experience in computational modeling and toxicity analysis for the deposition and dosimetry of toxins and chemicals and the mechanisms associated with e-cigarettes and vaping devices. In this project, we propose combining the expertise of the teams at Kitware and ARA to develop customized workflow for large data set storage and incorporating and analyzing machine learning techniques and results, respectively. We will demonstrate this effectiveness of the workflow using synthetic data generated using a computational framework of models. The specific aims of the Phase I project are: (1) Generate large datasets using high-fidelity computational modeling approaches; (2) Create an optimized workflow for ingesting large environmental health datasets for use in machine learning to calculate risk assessment; and (3) Develop a machine learning model to replace first principles models and predict risk assessment for environmental health.