(131) AI Literature Search to Support Rapid Analytics in Epidemiology

Publication Number: 1176

Background: Artificial Intelligence (AI) offers researchers an opportunity to assist with evidence generation that use the literature for their foundational data. Examples include targeted literature reviews (TLRs), systematic review, as well as answering specific questions, such as finding point-estimates for modeling from literature.

Objectives: To demonstrate two different TLR use cases where AI assisted in finding relevant literature in an efficient manner: #1- identifying all relevant studies, and #2 - extracting quantitative outcomes.

Methods: We leveraged an existing AI system that uses PubMed as a data source.[1] In case #1, we performed a TLR on recurrent wheezing, and used AI to identify retrospective, real-world studies and their ICD coding schemes. In case #2, we used AI to ascertain outcomes of third-line treatment for gastric cancer. The AI allowed our epidemiologists to search for and filter all relevant PubMed articles using AI-generated classifications of study methodology and AI extractions of results (e.g., interventions and outcomes).

Results: Starting with 316 articles from search in case #1, the AI enabled us to filter to the papers that measured “wheezing” as an outcome, automatically separate retrospective from prospective studies, and detect real-world studies as separate from other study types. This yielded 159 articles for screening, automatically eliminating nearly 50% of the irrelevant articles that failed the inclusion criteria. For case #2, the search returned more than 600 papers, from which the AI allowed us to narrow down to 17 papers with results focusing on third-line therapies. In this case, we had a manually performed search to compare against and found that the majority of the articles were relevant, comparable to the previous experiments on this AI.1 Overall, we found relevant articles and categorized pertinent information efficiently with less time resource allocation compared with manual review and extraction.

Conclusions: AI can help epidemiologists efficiently find useful articles to support rapid analytical workflows. Filtering based on AI-extracted results (e.g., outcomes or interventions) and methodology (e.g., study type) eliminates the need to review irrelevant articles. Future improvements of the AI platform could facilitate inclusion of (1) data sources outside of PubMed without extra steps to ingest the data (versus a human that can read anything), and (2) the ability to process languages other than English.

References:
1. Michelson M, Chow T, Martin NA, Ross M, Tee Qiao Ying A, Minton S
Artificial Intelligence for Rapid Meta-Analysis: Case Study on Ocular Toxicity of Hydroxychloroquine. J Med Internet Res 2020;22(8):e20007