Limits of search filter development.

OBJECTIVE: The research attempted to develop search filters for biomedical literature databases that improve retrieval of studies of clinical relevance for the nursing and rehabilitation professions. METHODS: Diagnostic testing framework compared machine-culled and practitioner-nominated search terms with a hand-tagged clinical literature database. RESULTS: We were unable to: (1) develop filters for nursing, likely because of the overlapping and expanding scope of practice for nurses in comparison with medical professionals, or (2) develop filters for rehabilitation, because of its broad scope and the profession's multifaceted understanding of "health and ability." CONCLUSIONS: We found limitations on search filter development for these health professions: nursing and rehabilitation.

Technology-mediated interventions for enhancing medication adherence.

BACKGROUND: Despite effective therapies for many conditions, patients find it difficult to adhere to prescribed treatments. Technology-mediated interventions (TMIs) are increasingly being used with the hope of improving adherence. OBJECTIVE: To assess the effects of TMI, intended to enhance patient adherence to prescribed medications, on both medication adherence and clinical outcomes. METHODS: A secondary in-depth analysis was conducted of the subset of studies that utilized technology in at least one component of the intervention from an updated Cochrane review on all interventions for enhancing medication adherence. We included studies that clearly described an information and communication technology or medical device as the sole or major component of the adherence intervention. RESULTS: Thirty-eight studies were eligible for in-depth review. Only seven had a low risk of bias for study design features, primary adherence, and clinical outcomes. Eighteen studies used a TMI for education and/or counseling, 11 studies used a TMI for self-monitoring and/or feedback, and nine studies used electronic reminders. Studies used a variety of TMIs, with telephone the most common technology in use. Studies targeted a wide distribution of diseases and used a variety of adherence and clinical outcome measures. A minority targeted children and adolescents. Fourteen studies reported significant effects in both adherence and clinical outcome measures. CONCLUSIONS: This review provides evidence for the inconsistent effectiveness of TMI for medication adherence and clinical outcomes. These results must be interpreted with caution due to a lack of high-quality studies.

High-performance information search filters for CKD content in PubMed, Ovid MEDLINE, and EMBASE.

BACKGROUND: Finding relevant articles in large bibliographic databases such as PubMed, Ovid MEDLINE, and EMBASE to inform care and future research is challenging. Articles relevant to chronic kidney disease (CKD) are particularly difficult to find because they are often published under different terminology and are found across a wide range of journal types. STUDY DESIGN: We used computer automation within a diagnostic test assessment framework to develop and validate information search filters to identify CKD articles in large bibliographic databases. SETTING & PARTICIPANTS: 22,992 full-text articles in PubMed, Ovid MEDLINE, or EMBASE. INDEX TEST: 1,374,148 unique search filters. REFERENCE TEST: We established the reference standard of article relevance to CKD by manual review of all full-text articles using prespecified criteria to determine whether each article contained CKD content or not. We then assessed filter performance by calculating sensitivity, specificity, and positive predictive value for the retrieval of CKD articles. Filters with high sensitivity and specificity for the identification of CKD articles in the development phase (two-thirds of the sample) were then retested in the validation phase (remaining one-third of the sample). RESULTS: We developed and validated high-performance CKD search filters for each bibliographic database. Filters optimized for sensitivity reached at least 99% sensitivity, and filters optimized for specificity reached at least 97% specificity. The filters were complex; for example, one PubMed filter included more than 89 terms used in combination, including "chronic kidney disease," "renal insufficiency," and "renal fibrosis." In proof-of-concept searches, physicians found more articles relevant to the topic of CKD with the use of these filters. LIMITATIONS: As knowledge of the pathogenesis of CKD grows and definitions change, these filters will need to be updated to incorporate new terminology used to index relevant articles. CONCLUSIONS: PubMed, Ovid MEDLINE, and EMBASE can be filtered reliably for articles relevant to CKD. These high-performance information filters are now available online and can be used to better identify CKD content in large bibliographic databases.

Increasing the quantity and quality of searching for current best evidence to answer clinical questions: protocol and intervention design of the MacPLUS FS Factorial Randomized Controlled Trials.

BACKGROUND & AIMS: Finding current best evidence for clinical decisions remains challenging. With 3,000 new studies published every day, no single evidence-based resource provides all answers or is sufficiently updated. McMaster Premium LiteratUre Service--Federated Search (MacPLUS FS) addresses this issue by looking in multiple high quality resources simultaneously and displaying results in a one-page pyramid with the most clinically useful at the top. Yet, additional logistical and educational barriers need to be addressed to enhance point-of-care evidence retrieval. This trial seeks to test three innovative interventions, among clinicians registered to MacPLUS FS, to increase the quantity and quality of searching for current best evidence to answer clinical questions. METHODS & DESIGN: In a user-centered approach, we designed three interventions embedded in MacPLUS FS: (A) a web-based Clinical Question Recorder; (B) an Evidence Retrieval Coach composed of eight short educational videos; (C) an Audit, Feedback and Gamification approach to evidence retrieval, based on the allocation of 'badges' and 'reputation scores.' We will conduct a randomized factorial controlled trial among all the 904 eligible medical doctors currently registered to MacPLUS FS at the hospitals affiliated with McMaster University, Canada. Postgraduate trainees (n=429) and clinical faculty/staff (n=475) will be randomized to each of the three following interventions in a factorial design (AxBxC). Utilization will be continuously recorded through clinicians' accounts that track logins and usage, down to the level of individual keystrokes. The primary outcome is the rate of searches per month per user during the six months of follow-up. Secondary outcomes, measured through the validated Impact Assessment Method questionnaire, include: utility of answers found (meeting clinicians' information needs), use (application in practice), and perceived usefulness on patient outcomes. DISCUSSION: Built on effective models for the point-of-care teaching, these interventions approach evidence retrieval as a clinical skill. If effective, they may offer the opportunity to enhance it for a large audience, at low cost, providing better access to relevant evidence across many top EBM resources in parallel. TRIAL REGISTRATION: ClinicalTrials.Gov NCT02038439.

Adherence measurement and patient recruitment methods are poor in intervention trials to improve patient adherence.

OBJECTIVES: To develop a scale and survey the measurement of patient adherence and patient recruitment, and to explore how these methods impact the results in randomized controlled trials of interventions to improve patient adherence to medications. STUDY DESIGN: Analytic survey of a purposively selected sample of patient adherence intervention trials from a systematic review, assessing the quality of adherence measurement and patient recruitment methods. RESULTS: We identified 44 different measures of adherence, with qualities ranging from valid and objective to unreliable and subjective. The median overall quality of measures of adherence was 5 (interquartile range [IQR], 3; range, 0-9, 9 is high quality). The quality of the measures was associated with variation in the estimate of adherence (Spearman r = 0.66; 95% confidence interval: 0.39, 0.83). The median overall quality of patient recruitment methods was 2 (IQR, 1; maximum score 6, higher is better). There was no significant correlation between the power of the trial to detect an effect and the quality of the patient recruitment methods. CONCLUSION: Measurement and recruitment methods in adherence trials varied considerably, and most methods were of low quality. Adherence research could be advanced by using higher quality measures of adherence and better selection and baseline assessment of study participants.

High-performance information search filters for acute kidney injury content in PubMed, Ovid Medline and Embase.

BACKGROUND: We frequently fail to identify articles relevant to the subject of acute kidney injury (AKI) when searching the large bibliographic databases such as PubMed, Ovid Medline or Embase. To address this issue, we used computer automation to create information search filters to better identify articles relevant to AKI in these databases. METHODS: We first manually reviewed a sample of 22 992 full-text articles and used prespecified criteria to determine whether each article contained AKI content or not. In the development phase (two-thirds of the sample), we developed and tested the performance of >1.3-million unique filters. Filters with high sensitivity and high specificity for the identification of AKI articles were then retested in the validation phase (remaining third of the sample). RESULTS: We succeeded in developing and validating high-performance AKI search filters for each bibliographic database with sensitivities and specificities in excess of 90%. Filters optimized for sensitivity reached at least 97.2% sensitivity, and filters optimized for specificity reached at least 99.5% specificity. The filters were complex; for example one PubMed filter included >140 terms used in combination, including 'acute kidney injury', 'tubular necrosis', 'azotemia' and 'ischemic injury'. In proof-of-concept searches, physicians found more articles relevant to topics in AKI with the use of the filters. CONCLUSIONS: PubMed, Ovid Medline and Embase can be filtered for articles relevant to AKI in a reliable manner. These high-performance information filters are now available online and can be used to better identify AKI content in large bibliographic databases.

Net improvement of correct answers to therapy questions after pubmed searches: pre/post comparison.

BACKGROUND: Clinicians search PubMed for answers to clinical questions although it is time consuming and not always successful. OBJECTIVE: To determine if PubMed used with its Clinical Queries feature to filter results based on study quality would improve search success (more correct answers to clinical questions related to therapy). METHODS: We invited 528 primary care physicians to participate, 143 (27.1%) consented, and 111 (21.0% of the total and 77.6% of those who consented) completed the study. Participants answered 14 yes/no therapy questions and were given 4 of these (2 originally answered correctly and 2 originally answered incorrectly) to search using either the PubMed main screen or PubMed Clinical Queries narrow therapy filter via a purpose-built system with identical search screens. Participants also picked 3 of the first 20 retrieved citations that best addressed each question. They were then asked to re-answer the original 14 questions. RESULTS: We found no statistically significant differences in the rates of correct or incorrect answers using the PubMed main screen or PubMed Clinical Queries. The rate of correct answers increased from 50.0% to 61.4% (95% CI 55.0%-67.8%) for the PubMed main screen searches and from 50.0% to 59.1% (95% CI 52.6%-65.6%) for Clinical Queries searches. These net absolute increases of 11.4% and 9.1%, respectively, included previously correct answers changing to incorrect at a rate of 9.5% (95% CI 5.6%-13.4%) for PubMed main screen searches and 9.1% (95% CI 5.3%-12.9%) for Clinical Queries searches, combined with increases in the rate of being correct of 20.5% (95% CI 15.2%-25.8%) for PubMed main screen searches and 17.7% (95% CI 12.7%-22.7%) for Clinical Queries searches. CONCLUSIONS: PubMed can assist clinicians answering clinical questions with an approximately 10% absolute rate of improvement in correct answers. This small increase includes more correct answers partially offset by a decrease in previously correct answers.

WhatisKT wiki: a case study of a platform for knowledge translation terms and definitions--descriptive analysis.

BACKGROUND: More than a hundred terms, often with unclear definitions and varying emphases, are used by health research and practice communities across the world who are interested in getting the best possible evidence applied (e.g., knowledge translation, implementation science, diffusion of innovations, and technology transfer). This makes finding published evidence difficult and can result in reduced, misinterpreted, or challenging interactions among professionals. Open dialogue and interaction among various professionals is needed to achieve consolidation of vocabulary. We use case report methods to describe how we sought to build an online tool to present the range of terms and facilitate the dialogue process across groups and disciplines interested in harnessing research evidence for healthcare. METHODS: We used a wiki platform from Wikispaces to present the problem of terminology and make a case and opportunity for collaboration on usage. Wikis are web sites where communities of users can collaborate online to build content and discuss progress. We gathered terms related to getting research into practice, sought published definitions, and posted these on the wiki (WhatisKT http://whatiskt.wikispaces.com/). We built the wiki in mid-2008 and promoted it through various groups and publications. This report describes the content of the site, our promotion efforts, use of the site, and how the site was used for collaboration up to the end of 2011. RESULTS: The WhatisKT wiki site now includes more than 120 pages. Traffic to the site has increased substantially from an average of 200 monthly visits in 2008 to 1700 in 2011. Visitors from 143 countries viewed the wiki in 2011, compared with 12 countries in 2008. However, most use has been limited to short term accesses of about 40 seconds per visit, and discussion of consolidation and solidifying terminology is conspicuously absent. CONCLUSIONS: Although considerable interest exists in the terms and definitions related to getting research into practice based on increasing numbers of accesses, use of the WhatisKT wiki site for anything beyond quick lookups was minimal. Additional efforts must be directed towards increasing the level of interaction among the members of the site to encourage collaboration on term use.

MEDLINE clinical queries are robust when searching in recent publishing years.

OBJECTIVE: To determine if the PubMed and Ovid MEDLINE clinical queries (which were developed in the publishing year 2000, for the purpose categories therapy, diagnosis, prognosis, etiology, and clinical prediction guides) perform as well when searching in current publishing years. METHODS: A gold standard database of recently published research literature was created using the McMaster health knowledge refinery (http://hiru.mcmaster.ca/hiru/HIRU_McMaster_HKR.aspx) and its continuously updated database, McMaster PLUS (http://hiru.mcmaster.ca/hiru/HIRU_McMaster_PLUS_projects.aspx). This database contains articles from over 120 clinical journals that are tagged for meeting or not meeting criteria for scientific merit and clinical relevance. The clinical queries sensitive ('broad') and specific ('narrow') search filters were tested in this gold standard database, and sensitivity and specificity were calculated and compared with those originally reported for the clinical queries. RESULTS: In all cases, the sensitivity of the highly sensitive search filters and the specificity of the highly specific search filters did not differ substantively when comparing results derived in 2000 with those derived in a more current database. In addition, in all cases, the specificities for the highly sensitive search filters and the sensitivities for the highly specific search filters remained above 50% when testing them in the current database. DISCUSSION: These results are reassuring for modern-day searchers. The clinical queries that were derived in the year 2000 perform equally well a decade later. CONCLUSION: The PubMed and Ovid MEDLINE clinical queries have been revalidated and remain a useful public resource for searching the world's medical literature for research that is most relevant to clinical care.

How current are leading evidence-based medical textbooks? An analytic survey of four online textbooks.

BACKGROUND: The consistency of treatment recommendations of evidence-based medical textbooks with more recently published evidence has not been investigated to date. Inconsistencies could affect the quality of medical care. OBJECTIVE: To determine the frequency with which topics in leading online evidence-based medical textbooks report treatment recommendations consistent with more recently published research evidence. METHODS: Summarized treatment recommendations in 200 clinical topics (ie, disease states) covered in four evidence-based textbooks--UpToDate, Physicians' Information Education Resource (PIER), DynaMed, and Best Practice--were compared with articles identified in an evidence rating service (McMaster Premium Literature Service, PLUS) since the date of the most recent topic updates in each textbook. Textbook treatment recommendations were compared with article results to determine if the articles provided different, new conclusions. From these findings, the proportion of topics which potentially require updating in each textbook was calculated. RESULTS: 478 clinical topics were assessed for inclusion to find 200 topics that were addressed by all four textbooks. The proportion of topics for which there was 1 or more recently published articles found in PLUS with evidence that differed from the textbooks' treatment recommendations was 23% (95% CI 17-29%) for DynaMed, 52% (95% CI 45-59%) for UpToDate, 55% (95% CI 48-61%) for PIER, and 60% (95% CI 53-66%) for Best Practice (χ(2) (3)=65.3, P<.001). The time since the last update for each textbook averaged from 170 days (range 131-209) for DynaMed, to 488 days (range 423-554) for PIER (P<.001 across all textbooks). CONCLUSIONS: In online evidence-based textbooks, the proportion of topics with potentially outdated treatment recommendations varies substantially.

The quality, breadth, and timeliness of content updating vary substantially for 10 online medical texts: an analytic survey.

OBJECTIVE: To evaluate the quality of evidence reporting, breadth of coverage, and timeliness of content updating of 10 selected online medical texts. STUDY DESIGN AND SETTING: Each text was assessed for quality based on an 11-item scale, which included items related to editorial policy and updating, appraisal, and transparent incorporation of newly published clinical research and evidence-based guidelines. Breadth of coverage was determined by the percentage of 60 randomly selected International Classification of Diseases 10 (ICD-10) codes covered by each of the texts. The same 60 ICD-10 codes were used to obtain a sample of topic chapters for the assessment of timeliness of updates. RESULTS: Quality scores ranged from a high of 9 of 11 points (Clinical Evidence) to a low of 0 of 11 points (PEPID), with a mean score of 6.7. Breadth of coverage ranged from 83% of randomly selected topics covered (UpToDate) to 25% (Clinical Evidence), with 6 of 10 texts covering 60% or more; average coverage across all texts was 57%. Variability was also observed with regard to average time since last content update, ranging from 3.5 (DynaMed) to 29 months (First Consult), with an average time since update of 12.4 months. CONCLUSION: No single resource was ideal and those seeking answers to clinical questions are well-advised not to rely solely on a single point-of-care product.

Dialysis search filters for PubMed, Ovid MEDLINE, and Embase databases.

BACKGROUND AND OBJECTIVES: Physicians frequently search bibliographic databases, such as MEDLINE via PubMed, for best evidence for patient care. The objective of this study was to develop and test search filters to help physicians efficiently retrieve literature related to dialysis (hemodialysis or peritoneal dialysis) from all other articles indexed in PubMed, Ovid MEDLINE, and Embase. DESIGN, SETTING, PARTICIPANTS, & MEASUREMENTS: A diagnostic test assessment framework was used to develop and test robust dialysis filters. The reference standard was a manual review of the full texts of 22,992 articles from 39 journals to determine whether each article contained dialysis information. Next, 1,623,728 unique search filters were developed, and their ability to retrieve relevant articles was evaluated. RESULTS: The high-performance dialysis filters consisted of up to 65 search terms in combination. These terms included the words "dialy" (truncated), "uremic," "catheters," and "renal transplant wait list." These filters reached peak sensitivities of 98.6% and specificities of 98.5%. The filters' performance remained robust in an independent validation subset of articles. CONCLUSIONS: These empirically derived and validated high-performance search filters should enable physicians to effectively retrieve dialysis information from PubMed, Ovid MEDLINE, and Embase.

Glomerular disease search filters for Pubmed, Ovid Medline, and Embase: a development and validation study.

BACKGROUND: Tools to enhance physician searches of Medline and other bibliographic databases have potential to improve the application of new knowledge in patient care. This is particularly true for articles about glomerular disease, which are published across multiple disciplines and are often difficult to track down. Our objective was to develop and test search filters for PubMed, Ovid Medline, and Embase that allow physicians to search within a subset of the database to retrieve articles relevant to glomerular disease. METHODS: We used a diagnostic test assessment framework with development and validation phases. We read a total of 22,992 full text articles for relevance and assigned them to the development or validation set to define the reference standard. We then used combinations of search terms to develop 997,298 unique glomerular disease filters. Outcome measures for each filter included sensitivity, specificity, precision, and accuracy. We selected optimal sensitive and specific search filters for each database and applied them to the validation set to test performance. RESULTS: High performance filters achieved at least 93.8% sensitivity and specificity in the development set. Filters optimized for sensitivity reached at least 96.7% sensitivity and filters optimized for specificity reached at least 98.4% specificity. Performance of these filters was consistent in the validation set and similar among all three databases. CONCLUSIONS: PubMed, Ovid Medline, and Embase can be filtered for articles relevant to glomerular disease in a reliable manner. These filters can now be used to facilitate physician searching.

Search filters can find some but not all knowledge translation articles in MEDLINE: an analytic survey.

OBJECTIVE: Advances from health research are not well applied giving rise to over- and underuse of resources and inferior care. Knowledge translation (KT), actions and processes of getting research findings used in practice, can improve research application. The KT literature is difficult to find because of nonstandardized terminology, rapid evolution of the field, and it is spread across several domains. We created multiple search filters to retrieve KT articles from MEDLINE. STUDY DESIGN AND SETTING: Analytic survey using articles from 12 journals tagged as having KT content and also as describing a KT application or containing a KT theory. RESULTS: Of 2,594 articles, 579 were KT articles of which 201 were about KT applications and 152 about KT theory. Search filter sensitivity (retrieval efficiency) maximized at 83%-94% with specificity (no retrieval of irrelevant material) approximately 50%. Filter performances were enhanced with multiple terms, but these filters often had reduced specificity. Performance was higher for KT applications and KT theory articles. These filters can select KT material although many irrelevant articles also will be retrieved. CONCLUSION: KT search filters were developed and tested, with good sensitivity but suboptimal specificity. Further research must improve their performance.

Impact of PubMed search filters on the retrieval of evidence by physicians.

BACKGROUND: Physicians face challenges when searching PubMed for research evidence, and they may miss relevant articles while retrieving too many nonrelevant articles. We investigated whether the use of search filters in PubMed improves searching by physicians. METHODS: We asked a random sample of Canadian nephrologists to answer unique clinical questions derived from 100 systematic reviews of renal therapy. Physicians provided the search terms that they would type into PubMed to locate articles to answer these questions. We entered the physician-provided search terms into PubMed and applied two types of search filters alone or in combination: a methods-based filter designed to identify high-quality studies about treatment (clinical queries "therapy") and a topic-based filter designed to identify studies with renal content. We evaluated the comprehensiveness (proportion of relevant articles found) and efficiency (ratio of relevant to nonrelevant articles) of the filtered and nonfiltered searches. Primary studies included in the systematic reviews served as the reference standard for relevant articles. RESULTS: The average physician-provided search terms retrieved 46% of the relevant articles, while 6% of the retrieved articles were relevant (corrected) (the ratio of relevant to nonrelevant articles was 1:16). The use of both filters together produced a marked improvement in efficiency, resulting in a ratio of relevant to nonrelevant articles of 1:5 (16 percentage point improvement; 99% confidence interval 9% to 22%; p < 0.003) with no substantive change in comprehensiveness (44% of relevant articles found; p = 0.55). INTERPRETATION: The use of PubMed search filters improves the efficiency of physician searches. Improved search performance may enhance the transfer of research into practice and improve patient care.

How well are journal and clinical article characteristics associated with the journal impact factor? a retrospective cohort study.

OBJECTIVE: Journal impact factor (JIF) is often used as a measure of journal quality. A retrospective cohort study determined the ability of clinical article and journal characteristics, including appraisal measures collected at the time of publication, to predict subsequent JIFs. METHODS: Clinical research articles that passed methods quality criteria were included. Each article was rated for relevance and newsworthiness by 3 to 24 physicians from a panel of more than 4,000 practicing clinicians. The 1,267 articles (from 103 journals) were divided 60∶40 into derivation (760 articles) and validation sets (507 articles), representing 99 and 88 journals, respectively. A multiple regression model was produced determining the association of 10 journal and article measures with the 2007 JIF. RESULTS: Four of the 10 measures were significant in the regression model: number of authors, number of databases indexing the journal, proportion of articles passing methods criteria, and mean clinical newsworthiness scores. With the number of disciplines rating the article, the 5 variables accounted for 61% of the variation in JIF (R(2) = 0.607, 95% CI 0.444 to 0.706, P<0.001). CONCLUSION: For the clinical literature, measures of scientific quality and clinical newsworthiness available at the time of publication can predict JIFs with 60% accuracy.

Search filter precision can be improved by NOTing out irrelevant content.

BACKGROUND: Most methodologic search filters developed for use in large electronic databases such as MEDLINE have low precision. One method that has been proposed but not tested for improving precision is NOTing out irrelevant content. OBJECTIVE: To determine if search filter precision can be improved by NOTing out the text words and index terms assigned to those articles that are retrieved but are off-target. DESIGN: Analytic survey. METHODS: NOTing out unique terms in off-target articles and testing search filter performance in the Clinical Hedges Database. MAIN OUTCOME MEASURES: Sensitivity, specificity, precision and number needed to read (NNR). RESULTS: For all purpose categories (diagnosis, prognosis and etiology) except treatment and for all databases (MEDLINE, EMBASE, CINAHL and PsycINFO), constructing search filters that NOTed out irrelevant content resulted in substantive improvements in NNR (over four-fold for some purpose categories and databases). CONCLUSION: Search filter precision can be improved by NOTing out irrelevant content.

Development and validation of filters for the retrieval of studies of clinical examination from Medline.

BACKGROUND: Efficiently finding clinical examination studies--studies that quantify the value of symptoms and signs in the diagnosis of disease-is becoming increasingly difficult. Filters developed to retrieve studies of diagnosis from Medline lack specificity because they also retrieve large numbers of studies on the diagnostic value of imaging and laboratory tests. OBJECTIVE: The objective was to develop filters for retrieving clinical examination studies from Medline. METHODS: We developed filters in a training dataset and validated them in a testing database. We created the training database by hand searching 161 journals (n = 52,636 studies). We evaluated the recall and precision of 65 candidate single-term filters in identifying studies that reported the sensitivity and specificity of symptoms or signs in the training database. To identify best combinations of these search terms, we used recursive partitioning. The best-performing filters in the training database as well as 13 previously developed filters were evaluated in a testing database (n = 431,120 studies). We also examined the impact of examining reference lists of included articles on recall. RESULTS: In the training database, the single-term filters with the highest recall (95%) and the highest precision (8.4%) were diagnosis[subheading] and "medical history taking"[MeSH], respectively. The multiple-term filter developed using recursive partitioning (the RP filter) had a recall of 100% and a precision of 89% in the training database. In the testing database, the Haynes-2004-Sensitive filter (recall 98%, precision 0.13%) and the RP filter (recall 89%, precision 0.52%) showed the best performance. The recall of these two filters increased to 99% and 94% respectively with review of the reference lists of the included articles. CONCLUSIONS: Recursive partitioning appears to be a useful method of developing search filters. The empirical search filters proposed here can assist in the retrieval of clinical examination studies from Medline; however, because of the low precision of the search strategies, retrieving relevant studies remains challenging. Improving precision may require systematic changes in the tagging of articles by the National Library of Medicine.