Optimal literature search for systematic reviews in surgery.

Background: The aim of the present study was to determine empirically which electronic databases contribute best to a literature search in surgical systematic reviews.
Methods: For ten published systematic reviews, the systematic literature searches were repeated in the databases MEDLINE, Web of Science, CENTRAL, and EMBASE. On the basis of these reviews, a gold standard set of eligible articles was created. Recall (%), precision (%), unique contribution (%), and numbers needed to read (NNR) were calculated for each database, as well as for searches of citing references and of the reference lists of related systematic reviews (hand search).
Results: CENTRAL yielded the highest recall (88.4%) and precision (8.3%) for randomized controlled trials (RCT), MEDLINE for non-randomized studies (NRS; recall 92.6%, precision 5.2%). The most effective combination of two databases plus hand searching for RCT was MEDLINE/CENTRAL (98.6% recall, NNR 97). Adding EMBASE marginally increased the recall to 99.3%, but with an NNR of 152. For NRS, the most effective combination was MEDLINE/Web of Science (99.5% recall, NNR 60).
Conclusions: For surgical systematic reviews, the optimal literature search for RCT employs MEDLINE and CENTRAL. For surgical systematic reviews of NRS, Web of Science instead of CENTRAL should be searched. EMBASE does not contribute substantially to reviews with a surgical intervention.