Abstract
Background: The typical approach to literature identification involves two discrete and successive steps: (i) formulating a search strategy (i.e., a set of Boolean queries) and (ii) manually identifying the relevant citations in the corpus returned by the query. We have developed a literature identification system (Pythia) that combines the query formulation and citation screening steps and uses modern approaches for text encoding (dense text embeddings) to represent the text of the citations in a form that can be used by information retrieval and machine learning algorithms.Methods: Pythia incorporates a set of natural-language questions with machine-learning algorithms to rank all PubMed citations based on relevance. Pythia returns the 100 top-ranked citations for all questions combined. These 100 articles are exported, and a human screener adjudicates the relevance of each abstract and tags words that indicate relevance. The “curated” articles are then exploited by Pythia to refine the search and re-rank the abstracts, and a new set of 100 abstracts is exported and screened/tagged, until convergence (i.e., no other relevant abstracts are retrieved) or for a set number of iterations (batches). Pythia performance was assessed using seven systematic reviews (three prospectively and four retrospectively). Sensitivity, precision, and the number needed to read were calculated for each review. Results: The ability of Pythia to identify the relevant articles (sensitivity) varied across reviews from a low of 0.09 for a sleep apnea review to a high of 0.58 for a diverticulitis review. The number of abstracts that a reviewer had to read to find one relevant abstract (NNR) was lower than in the manually screened project in four reviews, higher in two, and had mixed results in one. The reviews that had greater overall sensitivity retrieved more relevant citations in early batches, but neither study design, study size, nor specific key question significantly affected retrieval across all reviews.Conclusions: Future research should explore ways to encode domain knowledge in query formulation, possibly by incorporating a "reasoning" aspect to Pythia to elicit more contextual information and leveraging ontologies and knowledge bases to better enrich the questions used in the search.