Routine development of objectively derived search strategies

The conceptual approach is recommended by the pertinent literature on the development of high-quality search strategies [2, 11, 12]. The key feature of this approach is the expertise of the searcher, that is, her or his knowledge of the database structure, the thesaurus and the research topic, as well as the clinicians' subject knowledge [6]. This means, for example, that when the search aims to retrieve literature on "rheumatoid arthritis," appropriate synonyms and related terms for the text word part of the strategy need to be identified. Different sources can help identify synonyms and related terms, for example, in medical dictionaries such as MedlinePlus or the entry terms of the MeSH (that is, medical subject heading) database. A similar procedure is used to identify controlled vocabulary. However, it remains unclear how to decide which terms to include in the search strategy. Furthermore, it is difficult, and might even be impossible, to tell when the strategy is completed. Several synonyms and related terms are conceivable in the above-described example, such as "juvenile rheumatoid arthritis," "Caplan syndrome," "Felty syndrome," "rheumatoid nodule," "Sjögren syndrome," "ankylosing spondylitis," "Still disease," "sicca syndrome," "Bechterew disease" and so on. The strategy becomes increasingly extensive but also more prone to error because more search queries are used, increasing the risk of spelling errors, logical operator errors, line number errors, truncation errors and so on. Another disadvantage of this approach is that the lack of criteria for selecting terms can lead to lengthy and often unproductive discussions among the research team.

In general, search filters are developed to search bibliographic databases efficiently, that is, to increase the number of relevant studies gathered while minimizing the number of irrelevant studies [13, 14]. Search filters "are typically created by identifying and combining search terms to retrieve records with a common feature" [14] (p. 356). Attempts have been made to create different levels of strategies to cater to different users and their differing information needs [6]. The filters can be derived subjectively (expert-informed), objectively (research-based) or a combination of the two, that is, the search filter is derived subjectively but validated against a gold standard [6, 14]. Information specialists use textual analysis software on a set of relevant references to identify representative terms in this set [3, 6, 15, 16]. These empirically derived filters are then tested against a set of relevant and irrelevant records derived from a hand-search of SRs. There is general agreement that, whenever possible, objectively derived filters should be used.

Bak et al.[17] referred to Egger et al.[18] and stated that subjectively derived filters "draw their legitimacy from the expert knowledge ... and are therefore susceptible to the same criticisms as other reports of expert opinion" and that "as in standard biomedical evidence hierarchies, unvalidated filters based on expert opinion can be considered methodologically weak" [17] (p. 212).

On the basis of the example of brachytherapy for patients with prostate cancer, our aim in this paper is to describe the empirically guided development process for search strategies conducted by the German Institute for Quality and Efficiency in Health Care (Institut für Qualität und Wirtschaftlichkeit im Gesundheitswesen, or "IQWiG"). The paper is targeted mainly toward information specialists but may also provide useful information for other researchers with a specific interest in the development and validation of search strategies.

HTA agencies and other institutions that regularly conduct SRs require robust and reliable search strategies. In practice, IQWiG uses the described method for various areas and study designs, for example, for clinical and health economic topics as well as for RCTs and observational studies.

Ideally, the quality of developed search strategies should be as high as that of methodological or topic-specific filters. IQWiG therefore applies a predefined approach to the development and validation of search strategies for SRs, which is outlined in its General Methods paper (version 4.0) [23]. This approach is used for all elements of a search strategy that cannot be based on a tested search filter and usually refers to the content part of the search strategy (health conditions and interventions). The process of the development and validation of search strategies for SRs consists of four steps: (1) generation of a test set, (2) development of the search strategy (objectively derived approach), (3) validation of the search strategy and (4) standardized documentation.

To be able to develop and test a search strategy, a test set of relevant references is derived from SRs. For each HTA report, the information specialist conducts a preliminary search of the Cochrane Library (Cochrane Database of Systematic Reviews, Database of Abstracts of Reviews of Effects, and the Health Technology Assessment Database) to identify previous SRs in the area of interest. Because the Cochrane Collaboration specifies strict methodological standards for the preparation of SRs, it is a particularly trustworthy source for identifying this type of publication. When SRs on a similar research question are available and the search process, as well as the documented search strategy, is considered to be comprehensive, references included in the SRs are extracted to build the test set.

If SRs are not available, a precise strategy is developed and relevant articles are screened and selected by the review authors. For PubMed, the filter "Therapy/Narrow[filter]," which is accessed via the PubMed interface [24], is used for the precise search. In EMBASE, the precise filter "high specificity strategies" developed by McMaster University's Health Information Research Unit [25] can be accessed via Ovid.

The references identified in the SRs or the precise search are considered to be a "quasi-gold standard" [26]. The references identified are split randomly, using two-thirds for the development (development set) and one-third for the validation (validation set) of the search strategy.

After building the development set from the test set and importing the references into Endnote, a term frequency analysis is conducted using the Text Mining Package [27] of the R statistical software package [28]. On the basis of information derived from the titles and abstracts of the downloaded references, terms are ranked by frequency. Terms that are present in at least 20% of the references in the development set are selected for further examination. However, this ranking does not necessarily differentiate terms that are relevant to the research question from irrelevant terms in the target database. Therefore, a so-called population set consisting of a random sample of references is downloaded from the target database (for example, MEDLINE). This population set represents all references from the reference database and is compared to the development set. The most overrepresented terms related to the research question are used to develop the text word part of the search strategy. "Overrepresented" refers to the most frequent terms in the development set with a low sensitivity of 2% or less among the references in the population set [29]. The aim of this process is to identify those terms that are sensitive to the target references, but not to all references in the database.

Because of technical constraints, a simplified approach is adopted to identify controlled vocabulary. Terms are selected on the basis of their frequency in the development set and their relevance to the research question. For this purpose, tools such as PubMed PubReMiner, a free web service for searches in MEDLINE [30], or Endnote^®, a reference management software are used. In PubReMiner subheadings should be used with caution: Because controlled vocabulary is listed individually as soon as different subheadings are used, they need to be summarized first. Only then is it possible to check how often controlled vocabulary actually appears in the articles.

The process described above identifies effective candidate terms: text terms and controlled vocabulary that might be suitable for inclusion in the search strategy. The candidate terms are allocated to three main sets of terms according to the definitions in the Cochrane Handbook [2]: (1) terms used to search for the health condition of interest, (2) terms entered to search for the intervention evaluated and (3) terms used to search for the types of study design to be included (validated search filters can usually be applied here [25, 31‐35]).

The next step in assembling these terms in the actual search is undertaken manually in an iterative trial-and-error approach. Because SRs usually aim to apply highly sensitive search strategies, the strategy should capture all references from the development set with sufficient precision to prevent the retrieval of too many irrelevant references. During the course of an IQWiG project, the search strategy may be adjusted in consultation with the project team: for example, if a high sensitivity results in an excessive number of hits, a more precise strategy may be required. The results of the textual analysis are drawn upon to enable an informed and transparent decision regarding a change in strategy.

To confirm that the strategy developed works with a different set of references, the strategy is tested against a validation set. The validation set is also derived from SRs but contains different references than the development set. The strategy needs to be validated in the database for which the strategy was designed. The developed strategy is run in each database and compared to the validation set from that database using their accession numbers (for example, PMIDs in PubMed).

To ensure transparency, each step of the process needs to be documented. This includes documentation of the preliminary or the precise search strategy, the SRs and relevant references used for the development of the search strategy, and frequency tables, including terms and controlled vocabulary. This comprehensive internal documentation can also be used to discuss search strategies and for quality assurance purposes.

We performed a search for SRs in the Cochrane Library (Table 1). Three SRs on brachytherapy in patients with prostate cancer were eligible publications, from which 38 relevant references were extracted for the generation of the test set. After random separation of the test set, 25 references were available for the development set and 13 were available for the validation of the strategy (see Figure 1).

The analysis of text words of the 25 references from the development set, using the Text Mining Package in R, resulted in the generation of the list of frequencies in the development and population sets presented in Table 2. For example, the term "brachytherapy" was identified 19 times. As the development set included 25 references, this resulted in a sensitivity of 76%. A similar approach was chosen in the analysis of controlled vocabulary (listing them only according to frequency; see Table 3). For example, the term "brachytherapy" appeared a total of 20 times in the 25 references. A textual analysis was dispensed with for the study type, as validated study filters were available.

Taking the relevance of the topic into account, candidate terms were extracted from both lists and displayed in a new list. These candidate terms were allocated to one of three categories: health condition, intervention and "questionable terms" (terms for which it was unclear whether they should be considered in the strategy as well as terms that required further assessment) (see Table 4 and Additional material). In this context, it should be noted that "questionable terms" may also include terms that do not directly represent the intervention or health condition of interest. In our example, "Gleason" is a score for histologic grading. Such a term needs to be clarified a priori. The inclusion in the search strategy would be considered only if the specific terms did not identify the references from the test set in the categories "health condition" and "intervention."

The development of the search strategy was based on a trial-and-error approach whereby the candidate terms identified were entered into the bibliographic database with the corresponding syntax and we tested whether references from the development set could be detected (see line 8 of the search strategy in Table 5). In the example presented, the 25 hits of the development set were identified with the search strategy, meaning that sensitivity reached 100% (line 10 of the search strategy in Table 5). This means that there was no need to use questionable terms and that the search strategy could then be tested by means of the validation set.

The last step comprised the validation of the developed search strategy. For this purpose, the 13 references previously identified from the validation set were used. All references from the validation set could be identified (see line 10 of the search strategy in Table 6).

During the development and validation process, the following documents were stored for later quality control: the three SRs from which the test set was generated [36‐38]; the frequency tables, that is, the results of the textual analysis (Tables 2 and 3); the extraction of the candidate terms (Table 4); the prefinal search strategy (Table 5); and the validation results (Table 6).

Objectively derived and validated search strategies are an essential contribution to the development of high-quality search strategies. Some questions remain unanswered, however, and need to be addressed in future research. For instance, it is unclear how to handle situations where SRs are lacking or fail to fully cover the topic of interest. One approach could be to combine the concepts of interest from different SRs. For example, if the use of positron emission tomography (PET) in patients with gliomas is to be investigated, it might be appropriate to generate relevant references for this intervention, for example, from an SR on PET in patients with lymphoma, head and neck cancer and so on, and also to consider another SR that, for example, investigates the use of chemotherapy in patients with gliomas. This approach would ensure that a sufficient number of references would be retrieved to develop and validate single parts of the search strategy.

Another critical issue in the development and validation process is to determine the optimal number of references. So far, our experience shows that the suggested approach to develop the search strategy can even be used with a small sample of references. However, future research should explore sample size requirements for the development and validation process.

At IQWiG, we currently still use an iterative and essentially subjective approach to building the actual search strategy, which could be viewed as a limitation. Statistical approaches such as logistic regression or factor analysis [16] might be ways to find a more objective approach to performing this step. Further research is needed to determine whether these techniques produce competitive search strategies within an acceptable time frame.