Data Gathering Methods for Existing Chemicals Risk Evaluations under the Amended TSCA: Proposal for Using Natural Language Processing to Prioritize the Literature Search and Review

摘要

The amended Toxic Substances Control Act (TSCA) requires EPA to assess risks for existing chemicals. The risk evaluation must "integrate and assess available information on hazards and exposure for the conditions of use of the chemical substance, including information on specific risks of injury to health or the environment" and "describe the weight of scientific evidence for the identified hazards and exposure." To meet these requirements, EPA searched for all relevant data for the major components of an evaluation (engineering, exposure, fate, ecological hazard, and human health hazard), which produced thousands of results per chemical requiring manual review. In this presentation, we explore natural language processing (NLP), to automate discovery of relevant data from the peer reviewed literature. We describe the initial data gathering method for the first 10 chemicals undergoing risk evaluation focusing on the strategy for the exposure, engineering, and fate disciplines. We then explain how NLP can sort references into clusters using text patterns to prioritize extensive search results. We used a supervised form of clustering by adding a seed set (i.e., references identified a priori as relevant to the chemical) to the corpus to be clustered. Using NLP techniques, we compared the clustered results from the peer reviewed literature to the results from manually screening the same peer-reviewed references. Specifically, we calculated recall and elimination rate using clustering predictions compared to the entirely manual approach of reviewing all results. This approach requires a seed set which might not be easily obtained for future chemicals that have fewer known relevant references available. To address this, we also present the results using a generic, chemical-agnostic seed set that could be applied to future chemical assessments. The views expressed in this abstract are those of the authors and do not represent Agency policy or endorsement.