[Skip to Content]
Visit us on Facebook Visit us on FacebookVisit us on Linked In Visit us on Linked InVisit us on Twitter Visit us on TwitterVisit us on Facebook Visit us on InstagramVisit our RSS Feed View our RSS Feed
Region 5 Blog March 28th, 2024
CategoriesCategoriesCategories Contact UsContact Us ArchivesArchives Region/OfficeRegion SearchSearch

May

29

Date prong graphic

MeSH On Demand

Posted by on May 29th, 2014 Posted in: News from NLM, Technology


MeSH on Demand is a new tool announced in this month’s NLM Technical Bulletin and is available online for use: http://ii.nlm.nih.gov/Interactive/MeSHonDemand.shtml. This is one of the Natural Language Processing tools being developed in the Cognitive Science Branch of the Lister Hill National Center for Biomedical Communications, a division of the NLM. The on Demand tool analyzes chunks of text (up to 10000 characters) and identifies potentially related MeSH terms. From the MeSH on Demand page a user simply pastes in a piece of text, hits the “Find MeSH Terms” button, and a new page will be generated with suggested MeSH terms listed below the inputted text. According to the Technical Bulletin article, the tool will find “MeSH Headings, Publication Types, and Supplementary Concepts, but not Qualifiers (Subheadings).”

A disclaimer appears on the tool’s page that the results are generated via an automated, machine logic driven system which is meant to emulate human indexer thought. One can deduce from the disclaimer that we shouldn’t expect the underlying algorithms to understand all of the same textual nuances that a seasoned indexer would and it notes that “results will undoubtedly differ from any human-generated indexing.” This got me wondering though about how much the tool’s generated terms would differ from human-generated ones. To evaluate, I pasted in an abstract from an article on Computerized Provider Order Entry systems causing medication errors.  This was by no means meant as a methodical and thorough evaluation of MeSH on Demand.  Rather, this was simply meant to address personal curiosity and this particular article was selected using a “convenience sampling” technique (it was already open in a different tab).  This article had previously been indexed for MEDLINE with the following MeSH terms:

MeSH on Demand meanwhile came up with the following after processing the article abstract:

The aboutness of medication errors comes through in both sets of MeSH terms, but on Demand introduces a personnel focus (particularly male?), possibly drawing from the described audience of who took the administered questionnaire, whereas the MEDLINE indexers focused on the relevancy of the questionnaire itself.  Anti-bacterial agents can be chalked up to the tool only having access to the abstract of the article, while indexers who have scanned the whole article know that while the word antibacterial appears in the abstract, the antibiotic renewal notices are just a small piece of a bigger information system discussion (hence, Decision Support Systems, Clinical/standards* and Decision Support Systems, Clinical/standards*).

Users of MeSH on Demand aren’t limited to copying and pasting from the medical literature though.  I plugged in the first 10,000 characters of this New York Times article “All Circuits Are Busy” about a neuroscience researcher determined to map the brian’s complete neural network. The results from on Demand for this article are interesting not because of the MeSH terms themselves so much as the number of them it generated: 36, plus one supplementary concept (icodextrin).  Those familiar with MEDLINE subject indexing may be surprised at this result, as a typical indexed journal article will have 6-15 terms.  

Setting the tool’s disclaimer and discrepancies on the above test cases aside, one potential use for this tool is to identify possible MeSH terms for recently published articles of interest that are still listed as “in process” in PubMed.   An article I recently read that fits this bill is one from PLOS Computational Biology on using Wikipedia to detect flu trends (spoiler: Wikipedia article usage more accurately determined prevalence of flu-like illnesses than Google Flu trends). I pasted the abstract for this article into MeSH on Demand and these MeSH terms were suggested:

These terms would be able to give me a starting place for searching for related literature, and they happen to have a fair amount of overlap with the terms generated by a human indexer on a similar, older article about using Google Flu Trends to look for flu outbreaks:

While this tool certainly has some further development in its future, it is currently a viable tool for brainstorming MeSH terms to help build a PubMed query.

Image of the author ABOUT Mahria Lebow


Email author View all posts by
Developed resources reported in this program are supported by the National Library of Medicine (NLM), National Institutes of Health (NIH) under cooperative agreement number UG4LM012343 with the University of Washington.

NNLM and NETWORK OF THE NATIONAL LIBRARY OF MEDICINE are service marks of the US Department of Health and Human Services | Copyright | HHS Vulnerability Disclosure | Download PDF Reader