MEDLINE

According to the U.S. National Library of Medicine, "MEDLINE® (Medical Literature Analysis and Retrieval System Online) is the U.S. National Library of Medicine's® (NLM) premier bibliographic database that contains over 16 million references to journal articles in life sciences with a concentration on biomedicine. A distinctive feature of MEDLINE is that the records are indexed with NLM's Medical Subject Headings (MeSH®)."

PubMed is the National Library of Medicine's free online search system for MEDLINE.

PubMed provides feedback relevance with its "See related" feature.

Structure

MEDLINE® (Medical Literature Analysis and Retrieval System Online) is a database of predominantly biomedical bibliographic citations maintained by the U.S. National Library of Medicine (NLM). Each citation includes bibliographic data, abstract if available, links to full text of the article and keywords.

The process for selecting journals is described.

MeSH terms

The keywords are indexed with the NLM's Medical Subject Headings (MeSH®) and subheadings.

The important MeSH terms “Randomized Controlled Trial” and “Clinical Controlled Trial” were introduced in 1991 and 1995, respectively. The Cochrane Collaboration helps MEDLINE correctly retag articles with these terms.

The National Library of Medicine's Indexing Initiative is trying to automate assignment of MeSH terms. The National Library of Medicine is investigated whether indexing MeSH terms can be either fully or semi-automated. Indexing of MESH terms by human is assisted by the Medical Text Indexer (MTI).

There is some inconsistency in assignment of MeSH terms.

Methods to improve searching MEDLINE

Studies of searching MEDLINE.

Study	Setting	Method	Results	Comments
Tanaka 2011			A score based on MeSH terms, journal impact factor, and number of authors can predict future citation patterns
Bekhuis et al. 2010	Locating relevant studies for systematic reviews	Supervised machine learning with evolutionary SVM • Ensemble of four SVM classifiers (title;abstract;metadata)		Mean precision ranges from 26% to 37%
Wallace et al. 2010	Locating relevant studies for systematic reviews	Supervised machine learning with SVM • Ensemble of four SVM classifiers (title;abstract;MeSH;UMLS)		Reduced the number of articles to manually review by 40% to 50%.
Kilicoglu 2009	Identifying high quality studies of interventions (randomized controlled trials) in Internal Medicine	Supervised machine learning with SVM • Naïve Bayes • Boosting • Ensemble learning method (stacking)	82.5% precision 84.3% recall
Lin 2007	Oncoloogy Trained with “annotated bibliography of important literature on common problems in surgical oncology”(SSO-AB) Last updated 2001 (16% of SSO-AB are randomized controlled trials)			citation count per year (CCPY) outperformed citation count (CC) and journal impact factor (JIF)
Fu 2007				"impact factor and clinical query filters are unstable for different topics while a topic-specific impact factor and machine learning-based filter models appear more robust"
Cohen 2006	Locating relevant studies for updating systematic reviews
Aphinyanaphongs2006	Identifying high quality studies in Internal Medicine Trained with articles cited by ACP Journal Club	Supervised machine learning with SVM • Citation metrics		"machine learning filters outperform standard citation metrics...the filter models have to be built specifically for this task and gold standard...Previous research that claimed better performance of citation metrics than machine learning in one of the corpora examined here is attributed to using machine learning filters built for a different gold standard and task"
Bernstam 2006	Oncoloogy “annotated bibliography of important literature on common problems in surgical oncology”(SSO-AB) Last updated 2001 (16% of SSO-AB are randomized controlled trials) Trained with ACP Journal Club high quality studies	Supervised machine learning with SVM versus Journal impact factor, citation count per article, PageRank per article versus Pubmed's Clinical Queries (searches performed in 2004(?)	Precision of first 50 citations: PageRank 9% Citation count 8% Impact factor 2%	"Citation-based algorithms were more effective than noncitation-based algorithms"
Aphinyanaphongs 2005	Identifying high quality studies in Internal Medicine Trained with articles cited by ACP Journal Club	Supervised machine learning with SVM • Naïve Bayes • Boosting versus 1994 PubMed Clinical Queries	cell	Machine learning was more precise.
Notes: SVM. Support vector machine (see machine learning)

There is much ongoing research into improving MEDLINE search results.

Citation tracking

Citation tracking may help identify relevant studies in MEDLINE.

Clustering

Clustering search results may help.

Filters (hedges)

MEDLINE filters, also called hedges, are an optimal Boolean combination of search terms, both textword and MeSH terms, to search articles. Many filters have been made by the Hedges Team and are available as Clinical Queries at PubMed. Filters may improve efficieny (precision) of searches by physicians. The Clinical Queries at PubMed may improve the quality of articles retrieved.

Filters have been criticized for being imperfect.

Filters for article types

Evolution of search filters

Purpose category	Strategy with high sensitivity	Strategy with high specificity
1994 (developed with articles from 10 major journals)
Treatment	randomized controlled trial[Publication Type] OR drug therapy[MeSH Subheading] OR therapeutic use[MeSH Subheading] OR random*[Title/Abstract] • Sensitivity = 99% • Specificity = 74%	placebo*[Title/Abstract] OR (double[Title/Abstract] AND blind*[Title/Abstract]
Diagnosis
2005 (developed with articles from 160 journals)
Treatment	(clinical[Title/Abstract] AND trial[Title/Abstract]) OR clinical trials[MeSH Terms] OR clinical trial[Publication Type] OR random*[Title/Abstract] OR random allocation[MeSH Terms] OR therapeutic use[MeSH Subheading] • Sensitivity = 99% • Specificity = 70%	randomized controlled trial[Publication Type] OR (randomized[Title/Abstract] AND controlled[Title/Abstract] AND trial[Title/Abstract]) • Sensitivity = 93% • Specificity = 97%
Diagnosis	sensitiv*[Title/Abstract] OR sensitivity and specificity[MeSH Terms] OR diagnos*{Title/Abstract] OR diagnosis[MeSH:noexp] OR diagnostic * [MeSH:noexp] OR diagnosis,differential[MeSH:noexp] OR diagnosis[Subheading:noexp]	specificity[Title/Abstract]

One filter is for identifying randomized controlled trials. Many MEDLINE filters have been developed by the Hedges team supported by a grant from the National Library of Medicine. The filters were initially published in 1994 and then revised and published in 2005.

Examples include filters for randomized controlled trials and systematic reviews. Of note, the the filter for randomized controlled trials and retrieves systematic reviews very well.

Filters for studies of diagnostic test accuracy may or may not perform well. The reasons for missing studies may be due to incomplete indexing or articles by databases such as MEDLINE.

Filters for subject types

A filter have been developed for articles about kidney disease, dentistry, and about specific age ranges such as geriatrics.

Relevancy ranking

Although MEDLINE is usually searched for exact matches using Boolean terms, relevancy ranking has been studied. In an early comparison, relevancy ranking performed well; however, the Boolean version of MEDLINE did not fully use MeSH terms.

eTBLAST uses text mining to search for similar publications.

Citation analysis or PageRank

There are conflicting results over the role of ranking results based on citation counts or PageRank. A study using Google's own PageRank found PubMed's clinical queries to be better. However, a comparative study found better results for a metric analogous to PageRank for biomedical journals based on:

${\text{PageRank for the index article}}={\frac {\text{the number of articles citing the index article }}{\text{the number of articles cited by the index article}}}$

Machine learning

Machine learning methods in which the search engine seeks articles that more resemble the included articles, may be more accurate than Boolean methods (see EBMSearch below). However, the study by Aphinyanaphongs compared machine learning to the 1994 Boolean filters.

Machine learning may be improved by ensemble learning method using stacked generalization (or stacking) to emphasize the role of UMLS concepts and title words.

Machine learning may or may not be more accurate than citation based strategies. Citation or link strategies may improve upon text categorization.

Machine learning built for categorizing one gold standard may not work as well in another setting.

Research methods for comparative studies

In comparing the information retrieval of search strategies, there are two experimental methods.

1. If a complete test collection of articles is available that is already divided into articles of meeting inclusion criteria and articles that not meeting criteria, then each strategy is compared for its ability to successfully identify the articles meeting criteria (sensitivity) and to successfully exclude (specificity) the articles not meeting criteria. Sensitivity is also called "recall".
2. If a partial test collection is available that only consists of articles meeting inclusion criteria (for example, article meeting inclusion criteria for ACP Journal Club or articles included in a systematic review of a clinical topic or articles in an annotated bibliography), then the sensitivity is again the proportion of relevant articles identified by the strategy. However, the specificity is not computable. Instead, one of several related measures are calculated. These measures are all based on the positive predictive value (PPV) of the strategy. Analogous to PPV used in diagnostic testing, the PPV directly correlates with the prevalence of relevant articles in the collection and thus is not stable across prevalences.
   1. Precision, also called efficiency, is "the proportion of retrieved articles that meet criteria" and thus is the same as the PPV.
   2. Number Needed to Read (NNR) is 1/precision and is "how many papers in a journal have to be read to find one of adequate clinical quality and relevance." Of note, the NNR has been proposed as a metric to help libraries to decide which journals to subscribe to.
   3. Hit curve "is the number of important articles among the first n results."
   4. 11-point precision recall graph is similar to a receiver operating characteristic curve

Methods to access MEDLINE

There are many third party interfaces to search MEDLINE such as OVID. The National Library of Medicine's own search interface is PubMed (http://pubmed.gov). The National Library of Medicine maintains a list of search engines at http://www.ncbi.nlm.nih.gov/CBBresearch/Lu/search/.

PubMed

PubMed (http://pubmed.gov) is the National Library of Medicine's own free Internet access to MEDLINE. PubMed has been freely available since 1997.

EBM Search

EBM Search (http://www.ahsl.arizona.edu/ebmsearch/) is a federated medical search engine.

EBMSearch

EBMSearch (http://ebmsearch.org/) maintains its own copy of MEDLINE and uses machine learning to rank articles.

eTBLAST

eTBLAST uses text mining to search for similar publications.

GoPubMed

GoPubMed (http://www.GoPubMed.org/) applies social networking to MEDLINE.

HubMed

HubMed (http://www.hubmed.org/) does not maintain its own copy of MEDLILNE, but rather uses PubMed's EUtils web service to retrieve MEDLINE records stored at PubMed.

Medline Ranker

Medline Ranker uses machine learning.

MScanner

MScanner uses machine learning with Naïve Bayes classifier with two feature spaces (Medical Subject Headings (MeSH) and the journal of publication).

Ovid

• Ovid Searching Tips

SUMSearch

SUMSearch (http://sumsearch.org/) is a federated medical search engine. It does not maintain its own copy of MEDLINE, but rather queries PubMed and revises searches too few or too many citations are retrieved. At the same time, SUMSearch queries the National Guidelines Clearinghouse and other resources.