Prediction of relevant biomedical documents: a human microbiome case study

Paul Thompson, Juliette C. Madan, Jason H. Moore

Retrieving relevant biomedical literature has become increasingly difficult due to the large volume and rapid growth of biomedical publication. A query to a biomedical retrieval system often retrieves hundreds of results. Since the searcher will not likely consider all of these documents, ranking the documents is important. Ranking by recency, as PubMed does, takes into account only one factor indicating potential relevance. This study explores the use of the searcher's relevance feedback judgments to support relevance ranking based on features more general than recency. It was found that the researcher's relevance judgments could be used to accurately predict the relevance of additional documents: both using tenfold cross-validation and by training on publications from 2008-2010 and testing on documents from 2011. This case study has shown the promise for relevance feedback to improve biomedical document retrieval. A researcher's judgments as to which initially retrieved documents are relevant, or not, can be leveraged to predict additional relevant documents.