Using Artificial Neural Networks to Predict Disease Associations for Chemicals Present in Burn Pit Emissions

In June of 2015, 27,378 of the 28,000 returning Operation Iraqi Freedom/Operation Enduring Freedom (OIF/OEF) veterans report being exposed to burn pits. According to Barth et al. (2014), 9,660 returning OIF/OEF veterans were diagnosed with respiratory diseases, to include asthma, bronchitis, and sinusitis, thus strengthening the need to develop decision support tools that can be used to understand the relationships between chemical exposure and disease. In this study an Artificial Neural Network (ANN) was used to predict the chemical-disease associations for burn pit constituents. Ten burn pit constituents were tested using varying hidden layers, similar chemical structure relationships, and three Training, Validation, and Testing (TVT) ratios. The ANN predicted misidentification rates of 73% or greater when the hidden layer size varied between 1 and 5. Misidentification rates of 75% or greater were observed for ANN simulations when the TVT ratios ranged from 60/20/20 to 80/10/10. ANN-based screening of chemical groups containing chemicals with benzene rings and chemicals containing hydrocarbon chains produced misidentification rates of 73% or greater, and R2 values of 0.0762 and lower. Hidden Layer size, TVT ratios, and chemical structure had little effect on the model’s performance; additional training data is needed to improve the predictive capability of the ANN. The ANN-based screening of individual burn pit constituents produced several chemicals with R2 values greater than 0.8. These chemicals have been prioritized to further develop predictive ANN models for human health force support, resulting in the first research screening burn pit constituents with an ANN, and the first to prioritize burn pit emissions for future testing