|Home | About | Journals | Submit | Contact Us | Français|
Increasingly, clinicians are doing their own searches using large biomedical literature databases . Choice of database can influence the success of a search . MEDLINE is often searched first due to its free access through the PubMed interface and its broad coverage of the biomedical literature, including nursing, dentistry, paramedic professions, reproductive biology, and clinical and experimental medicine. EMBASE searching is not free but has greater coverage of European and non-English language publications and topics such as pharmaceuticals, psychiatry, toxicology, and alternative medicine. The overlap of EMBASE and MEDLINE is estimated to be 10% to 87%, depending on the topic .
To date, search strategy development has focused more on MEDLINE [4–9] than on EMBASE. Search strategies developed for MEDLINE cannot be directly translated for use in other databases because indexing practices vary and thesaurus terms are not equivalent across databases.
In the 1990s, the Hedges Team developed MEDLINE search strategies for a small subset of 10 journals [10, 11]. This work has been expanded using data from 161 journals indexed in MEDLINE in 2000 and a 55-journal subset for EMBASE. The MEDLINE strategies [12–18] and some EMBASE strategies [19–22] have previously been published. This report compares the sensitivity and specificity of top-performing search strategies for detecting treatment and review articles in MEDLINE and EMBASE.
The methods used in this study have been detailed elsewhere . Briefly, the operating characteristics of search strategies were compared with a manual review of 161 health care journals in 2000 for MEDLINE  and a 55-journal subset for EMBASE . Six research assistants manually assessed all articles for studies meeting methodologic criteria in 7 purpose categories (treatment, causation, prognosis, diagnosis, economics, clinical prediction, and reviews). The authors reported purpose category definitions and methodologic criteria in an earlier paper . To evaluate search strategies designed to retrieve studies meeting basic methodologic criteria, index terms and text-words related to research design features were run as search strategies. Search strategies were treated as “diagnostic tests” for sound studies, and the manual review (hand search) was treated as the “gold standard.” Operating characteristics of the search strategies were determined. Top-performing strategies for detecting sound treatment and systematic review articles in MEDLINE [17, 18] and EMBASE [26, 27] were compared.
In both MEDLINE and EMBASE, strategies used text-words and index terms. The high-sensitivity strategies in both databases performed similarly but used different combinations of terms except for the same text-word, “random:.tw.” The high-sensitivity MEDLINE strategy used the publication type, “clinical trial,” and the exploded therapeutic use subheading, “tu.xs,” neither of which was supported in EMBASE. The high-sensitivity EMBASE strategy used the exploded subject heading, “health care quality,” a term not supported in MEDLINE. Because several terms in these top-performing strategies were uniquely supported in their respective databases, it was not possible to directly test them in the opposite database.
Strategies that yielded high specificity and minimized the difference between sensitivity and specificity performed slightly better overall in MEDLINE than in EMBASE. MEDLINE strategies used the publication type, “randomized controlled trial,” which EMBASE did not support.
Strategies in both MEDLINE and EMBASE used text-words and index terms. The high-sensitivity strategy in MEDLINE was more sensitive than the comparable EMBASE strategy but had lower specificity. Both strategies included the publication type, “review,” but the MEDLINE strategy included an additional publication type, “meta analysis,” which was not supported in EMBASE. The EMBASE strategy used the subject heading, “methodology,” which was not supported in MEDLINE. In both databases, the textword, “search:.,” was a top-performing term.
The most specific strategies in MEDLINE and EMBASE were similarly specific, but the MEDLINE strategy had better sensitivity. The MEDLINE and EMBASE strategies used similar text-words (“MEDLINE.tw.” and “systematic review.tw”). In MEDLINE, the high-specificity strategy used the publication type, “meta analysis,” and the journal name, “Cochrane Database of Systematic Reviews,” both of which were uniquely supported in MEDLINE. This latter term retrieves Cochrane reviews, all of which would meet the criteria for a sound review. EMBASE did not index this publication. In EMBASE, the single term, “meta analysis.sh.,” also yielded high specificity.
The top-performing strategy for minimizing the difference between sensitivity and specificity in MEDLINE outperformed the comparable EMBASE strategy. The strategies were similar in the use of the terms, “search:.tw.” and “review.pt.,” but MEDLINE used the additional publication type, “meta analysis.”
Top-performing filters in MEDLINE and EMBASE were not equivalent in the search terms used, although some overlap existed. MEDLINE search strategies generally outperformed EMBASE strategies, but, in both databases, strategies achieved high sensitivities and specificities, whereas precision peaked at about 50%. Precision is inevitably low in large multipurpose databases .
EMBASE incorporated fewer publication types than MEDLINE, and, in the strategies examined, the best MEDLINE strategies contained several publication types that were not supported in EMBASE. Table 3 displays the performance of these publication types, plus “review.pt” (supported in both databases), and compares them with the similar term searched as a subject heading or text-word in MEDLINE and EMBASE. All MEDLINE publication types attained specificities greater than 90% and reasonably high sensitivities (>77%), except the term, “meta analysis.pt” (possibly because the definition used for “reviews” encompassed reviews broader than only meta-analyses). EMBASE subject headings generally yielded better sensitivities than the similar text-words. Text-words had overall lower sensitivity but slightly higher specificity than index terms, a finding that is consistent with previous research .
As shown in Table 3, the higher sensitivities achieved by searching with publication types alone over text-words suggest an enhanced quality of indexing for publication types. This gives insight to the strategies in Tables 1 and and2,2, where use of publication types in MEDLINE yielded generally higher sensitivities. It follows that EMBASE strategies might be enhanced by expanding the available range of relevant publication types.
In the absence of a greater range of publication types in EMBASE, comprehensive searches should use subject headings with a methodologic focus over similar text-words to potentially improve sensitivity while maintaining a smaller difference between sensitivity and specificity. Text-word searching, however, remains valuable because text-words are versatile, do not rely on consistent indexing, and can be used to detect citation of “in process” articles that have not yet been assigned indexing terms. Comparing the use of text-words in the two databases is difficult because of differences in the spelling of key words, such as “randomized” in the United States and “randomised” in most European countries, or the different use of various methodologic concepts. An example of different use of methodologic concepts can be made with the term, “overview.” In the United States, an “overview” is considered to be a narrative review (a review that does not meet the methodologic criteria used in this study), whereas in the United Kingdom, an “overview” is considered to be a comprehensive systematic review (a review that would meet the methodologic criteria used in this study). Text-word searching requires the use of exact spelling of text-words in articles. Authors who use commonly understood text-words in titles and abstracts can facilitate a successful search.
Further work is necessary to determine the degree of overlap of citations retrieved using MEDLINE and EMBASE strategies. Better search strategies can probably be developed, but they would likely be more complex than those reported here, and improvements in sensitivity will be offset by decreases in specificity, or vice versa.
Optimal filters in MEDLINE and EMBASE for detecting treatment and review articles attained high performance using different search terms. Extra publication types in MEDLINE appeared to increase search sensitivity, and MEDLINE filters generally performed a little better than EMBASE filters. Considering the unique content coverage and search terms available in MEDLINE and EMBASE, information professionals would be best served by searching the database that is most relevant to their setting or searching both databases to improve a comprehensive search.
Haynes and Wilczynski prepared grant submissions in relation to this project, supplied intellectual content to the collection and analysis of data, and were involved in data analysis and staff supervision. Wilczynski and Wong participated in the data collection. All authors drafted, commented on, and approved the final manuscript.
The funding organization (National Library of Medicine, USA.) had no role in the design and conduct of the study; in the collection, analysis, and interpretation of data; or in the preparation, review, and approval of the manuscript. The Hedges Team includes Angela Eady, Brian Haynes, Susan Marks, Ann McKibbon, Doug Morgan, Cindy Walker-Dilks, Stephen Walter, Stephen Werre, Nancy Wilczynski, and Sharon Wong, all in the Health Information Research Unit, Department of Clinical Epidemiology and Biostatistics at McMaster University, Hamilton, Ontario, Canada.
*Source of funding: National Library of Medicine, grant no. 5R01LM06866-04.