Development of search strategies

At Erasmus MC, search strategies for systematic reviews are often designed via a librarian-mediated search service. The information specialists of Erasmus MC developed an efficient method that helps them perform searches in many databases in a much shorter time than other methods. This method of literature searching and a pragmatic evaluation thereof are published in separate journal articles [21, 22]. In short, the method consists of an efficient way to combine thesaurus terms and title/abstract terms into a single line search strategy. This search is then optimized. Articles that are indexed with a set of identified thesaurus terms, but do not contain the current search terms in title or abstract, are screened to discover potential new terms. New candidate terms are added to the basic search and evaluated. Once optimal recall is achieved, macros are used to translate the search syntaxes between databases, though manual adaptation of the thesaurus terms is still necessary.

Review projects at Erasmus MC cover a wide range of medical topics, from therapeutic effectiveness and diagnostic accuracy to ethics and public health. In general, searches are developed in MEDLINE in Ovid (Ovid MEDLINE® In-Process & Other Non-Indexed Citations, Ovid MEDLINE® Daily and Ovid MEDLINE®, from 1946); Embase.com (searching both Embase and MEDLINE records, with full coverage including Embase Classic); the Cochrane Central Register of Controlled Trials (CENTRAL) via the Wiley Interface; Web of Science Core Collection (hereafter called Web of Science); PubMed restricting to records in the subset “as supplied by publisher” to find references that not yet indexed in MEDLINE (using the syntax publisher [sb]); and Google Scholar. In general, we use the first 200 references as sorted in the relevance ranking of Google Scholar. When the number of references from other databases was low, we expected the total number of potential relevant references to be low. In this case, the number of hits from Google Scholar was limited to 100. When the overall number of hits was low, we additionally searched Scopus, and when appropriate for the topic, we included CINAHL (EBSCOhost), PsycINFO (Ovid), and SportDiscus (EBSCOhost) in our search.

Beginning in May 2013, the number of records retrieved from each search for each database was recorded at the moment of searching. The complete results from all databases used for each of the systematic reviews were imported into a unique EndNote library upon search completion and saved without deduplication for this research. The researchers that requested the search received a deduplicated EndNote file from which they selected the references relevant for inclusion in their systematic review. All searches in this study were developed and executed by W.M.B.

Determining relevant references of published reviews

We searched PubMed in July 2016 for all reviews published since 2014 where first authors were affiliated to Erasmus MC, Rotterdam, the Netherlands, and matched those with search registrations performed by the medical library of Erasmus MC. This search was used in earlier research [21]. Published reviews were included if the search strategies and results had been documented at the time of the last update and if, at minimum, the databases Embase, MEDLINE, Cochrane CENTRAL, Web of Science, and Google Scholar had been used in the review. From the published journal article, we extracted the list of final included references. We documented the department of the first author. To categorize the types of patient/population and intervention, we identified broad MeSH terms relating to the most important disease and intervention discussed in the article. We copied from the MeSH tree the top MeSH term directly below the disease category or, in to case of the intervention, directly below the therapeutics MeSH term. We selected the domain from a pre-defined set of broad domains, including therapy, etiology, epidemiology, diagnosis, management, and prognosis. Lastly, we checked whether the reviews described limiting their included references to a particular study design.

To identify whether our searches had found the included references, and if so, from which database(s) that citation was retrieved, each included reference was located in the original corresponding EndNote library using the first author name combined with the publication year as a search term for each specific relevant publication. If this resulted in extraneous results, the search was subsequently limited using a distinct part of the title or a second author name. Based on the record numbers of the search results in EndNote, we determined from which database these references came. If an included reference was not found in the EndNote file, we presumed the authors used an alternative method of identifying the reference (e.g., examining cited references, contacting prominent authors, or searching gray literature), and we did not include it in our analysis.

Data analysis

We determined the databases that contributed most to the reviews by the number of unique references retrieved by each database used in the reviews. Unique references were included articles that had been found by only one database search. Those databases that contributed the most unique included references were then considered candidate databases to determine the most optimal combination of databases in the further analyses.

Performance of a search can be expressed in different ways. Depending on the goal of the search, different measures may be optimized. In the case of a clinical question, precision is most important, as a practicing clinician does not have a lot of time to read through many articles in a clinical setting. When searching for a systematic review, recall is the most important aspect, as the researcher does not want to miss any relevant references. As our research is performed on systematic reviews, the main performance measure is recall.

We identified all included references that were uniquely identified by a single database. For the databases that retrieved the most unique included references, we calculated the number of references retrieved (after deduplication) and the number of included references that had been retrieved by all possible combinations of these databases, in total and per review. For all individual reviews, we determined the median recall, the minimum recall, and the percentage of reviews for which each single database or combination retrieved 100% recall.

For each review that we investigated, we determined what the recall was for all possible different database combinations of the most important databases. Based on these, we determined the percentage of reviews where that database combination had achieved 100% recall, more than 95%, more than 90%, and more than 80%. Based on the number of results per database both before and after deduplication as recorded at the time of searching, we calculated the ratio between the total number of results and the number of results for each database and combination.

Improvement of precision was calculated as the ratio between the original precision from the searches in all databases and the precision for each database and combination.

To compare our practice of database usage in systematic reviews against current practice as evidenced in the literature, we analyzed a set of 200 recent systematic reviews from PubMed. On 5 January 2017, we searched PubMed for articles with the phrase “systematic review” in the title. Starting with the most recent articles, we determined the databases searched either from the abstract or from the full text until we had data for 200 reviews. For the individual databases and combinations that were used in those reviews, we multiplied the frequency of occurrence in that set of 200 with the probability that the database or combination would lead to an acceptable recall (which we defined at 95%) that we had measured in our own data.

Unique references per database

Overall performance

Probability of appropriate recall

We calculated the recall for individual databases and databases in all possible combination for all reviews included in the research. Figure 1 shows the percentages of reviews where a certain database combination led to a certain recall. For example, in 48% of all systematic reviews, the combination of Embase and MEDLINE (with or without Cochrane CENTRAL; Cochrane CENTRAL did not add unique relevant references) reaches a recall of at least 95%. In 72% of studied systematic reviews, the combination of Embase, MEDLINE, Web of Science, and Google Scholar retrieved all included references. In the top bar, we present the results of the complete database searches relative to the total number of included references. This shows that many database searches missed relevant references.

Differences between domains of reviews

We analyzed whether the added value of Web of Science and Google Scholar was dependent of the domain of the review. For 55 reviews, we determined the domain. See Fig. 2 for the comparison of the recall of Embase, MEDLINE, and Cochrane CENTRAL per review for all identified domains. For all but one domain, the traditional combination of Embase, MEDLINE, and Cochrane CENTRAL did not retrieve enough included references. For four out of five systematic reviews that limited to randomized controlled trials (RCTs) only, the traditional combination retrieved 100% of all included references. However, for one review of this domain, the recall was 82%. Of the 11 references included in this review, one was found only in Google Scholar and one only in Web of Science.

Reduction in number of results

We calculated the ratio between the number of results found when searching all databases, including databases not included in our analyses, such as Scopus, PsycINFO, and CINAHL, and the number of results found searching a selection of databases. See Fig. 3 for the legend of the plots in Figs. 4 and 5. Figure 4 shows the distribution of this value for individual reviews. The database combinations with the highest recall did not reduce the total number of results by large margins. Moreover, in combinations where the number of results was greatly reduced, the recall of included references was lower.

Improvement of precision

To determine how searching multiple databases affected precision, we calculated for each combination the ratio between the original precision, observed when all databases were searched, and the precision calculated for different database combinations. Figure 5 shows the improvement of precision for 15 databases and database combinations. Because precision is defined as the number of relevant references divided by the number of total results, we see a strong correlation with the total number of results.

Status of current practice of database selection

Strengths and limitations

We did not investigate whether the loss of certain references had resulted in changes to the conclusion of the reviews. Of course, the loss of a minor non-randomized included study that follows the systematic review’s conclusions would not be as problematic as losing a major included randomized controlled trial with contradictory results. However, the wide range of scope, topic, and criteria between systematic reviews and their related review types make it very hard to answer this question.

We found that two databases previously not recommended as essential for systematic review searching, Web of Science and Google Scholar, were key to improving recall in the reviews we investigated. Because this is a novel finding, we cannot conclude whether it is due to our dataset or to a generalizable principle. It is likely that topical differences in systematic reviews may impact whether databases such as Web of Science and Google Scholar add value to the review. One explanation for our finding may be that if the research question is very specific, the topic of research might not always be mentioned in the title and/or abstract. In that case, Google Scholar might add value by searching the full text of articles. If the research question is more interdisciplinary, a broader science database such as Web of Science is likely to add value. The topics of the reviews studied here may simply have fallen into those categories, though the diversity of the included reviews may point to a more universal applicability.

Although we searched PubMed as supplied by publisher separately from MEDLINE in Ovid, we combined the included references of these databases into one measurement in our analysis. Until 2016, the most complete MEDLINE selection in Ovid still lacked the electronic publications that were already available in PubMed. These could be retrieved by searching PubMed with the subset as supplied by publisher. Since the introduction of the more complete MEDLINE collection Epub Ahead of Print, In-Process & Other Non-Indexed Citations, and Ovid MEDLINE®, the need to separately search PubMed as supplied by publisher has disappeared. According to our data, PubMed’s “as supplied by publisher” subset retrieved 12 unique included references, and it was the most important addition in terms of relevant references to the four major databases. It is therefore important to search MEDLINE including the “Epub Ahead of Print, In-Process, and Other Non-Indexed Citations” references.

These results may not be generalizable to other studies for other reasons. The skills and experience of the searcher are one of the most important aspects in the effectiveness of systematic review search strategies [23–25]. The searcher in the case of all 58 systematic reviews is an experienced biomedical information specialist. Though we suspect that searchers who are not information specialists or librarians would have a higher possibility of less well-constructed searches and searches with lower recall, even highly trained searchers differ in their approaches to searching. For this study, we searched to achieve as high a recall as possible, though our search strategies, like any other search strategy, still missed some relevant references because relevant terms had not been used in the search. We are not implying that a combined search of the four recommended databases will never result in relevant references being missed, rather that failure to search any one of these four databases will likely lead to relevant references being missed. Our experience in this study shows that additional efforts, such as hand searching, reference checking, and contacting key players, should be made to retrieve extra possible includes.

Based on our calculations made by looking at random systematic reviews in PubMed, we estimate that 60% of these reviews are likely to have missed more than 5% of relevant references only because of the combinations of databases that were used. That is with the generous assumption that the searches in those databases had been designed sensitively enough. Even when taking into account that many searchers consider the use of Scopus as a replacement of Embase, plus taking into account the large overlap of Scopus and Web of Science, this estimate remains similar. Also, while the Scopus and Web of Science assumptions we made might be true for coverage, they are likely very different when looking at recall, as Scopus does not allow the use of the full features of a thesaurus. We see that reviewers rarely use Web of Science and especially Google Scholar in their searches, though they retrieve a great deal of unique references in our reviews. Systematic review searchers should consider using these databases if they are available to them, and if their institution lacks availability, they should ask other institutes to cooperate on their systematic review searches.

The major strength of our paper is that it is the first large-scale study we know of to assess database performance for systematic reviews using prospectively collected data. Prior research on database importance for systematic reviews has looked primarily at whether included references could have theoretically been found in a certain database, but most have been unable to ascertain whether the researchers actually found the articles in those databases [10, 12, 16, 17, 26]. Whether a reference is available in a database is important, but whether the article can be found in a precise search with reasonable recall is not only impacted by the database’s coverage. Our experience has shown us that it is also impacted by the ability of the searcher, the accuracy of indexing of the database, and the complexity of terminology in a particular field. Because these studies based on retrospective analysis of database coverage do not account for the searchers’ abilities, the actual findings from the searches performed, and the indexing for particular articles, their conclusions lack immediate translatability into practice. This research goes beyond retrospectively assessed coverage to investigate real search performance in databases. Many of the articles reporting on previous research concluded that one database was able to retrieve most included references. Halladay et al. [10] and van Enst et al. [16] concluded that databases other than MEDLINE/PubMed did not change the outcomes of the review, while Rice et al. [17] found the added value of other databases only for newer, non-indexed references. In addition, Michaleff et al. [26] found that Cochrane CENTRAL included 95% of all RCTs included in the reviews investigated. Our conclusion that Web of Science and Google Scholar are needed for completeness has not been shared by previous research. Most of the previous studies did not include these two databases in their research.