Background: Systematic searches for Cochrane systematic reviews (CSRs) include some search terms derived from PICO (P = participants, I = intervention, C = control, O = outcome) and for the study design. Searches often produce thousands of potentially relevant references. Manually separating true positives from false positives is tedious and time-consuming.
Objectives: To see if search strategies for CSRs can be systematically improved in terms of efficiency.
Methods: As material for developing search strategies we used references that had been earlier identified as relevant for our own occupational health CSRs on (1) treatments for dysphonia (RCTs only); (2) preventing injuries in the construction industry (both RCTs and non-randomised studies). For both, we created a gold standard from only those references that could be found in PubMed. In the original searches we used both Medical Subject Heading (MeSH) terms and text word-terms describing P and O in PICO. For the study and intervention type elements of the searches, we used the revised Cochrane search strategy and the Cochrane Occupational Health Field (COHF) most sensitive search strategy. Based on the content of the retrieved articles we tried to formulate more efficient search strategies with more appropriate search terms both as text words and MeSH terms. As study type, the following strategies (filters), previously published, for both search term groups were used I) "Most specific COHF search strategy" [Verbeek et al]; II) "Most sensitive COHF search strategy" [Verbeek et al]; III) Revised Cochrane search strategy [Robinson et al]; IV) Clinical Queries: therapy, narrow [PubMed filter]; V) Clinical Queries: therapy, broad [PubMed filter]. Values for sensitivity and Number Needed to Read (NNR = 1/precision) were used for evaluation.
Results: Sensitivity could be considerably improved for the dysphonia review compared to the original search by using better text words and search string for RCTs only. For the construction industry review it was not possible to improve sensitivity but we could decrease the NNR. The use of MeSH terms decreases the sensitivity too much to warrant the increase in NNR and does not have added value in Cochrane reviews.
Conclusions: With the hindsight of the retrieved articles, the sensitivity and NNR of search strategy could be substantially improved for these two search topics. This was mostly due to the use of more appropriate search terms for the P and O in the PICO. The next challenge is to be able to make judgements beforehand.

Table 1. Search results for dysphonia review; number of retrieved articles = 43
SEARCH STRATEGY sensitivity (%) NNR
Original search for CSR 49 76
Most specific, tw 9 10
Most sensitive, tw 47 22
Revised Cochrane, tw 77 58
Clinical Query narrow, tw 23 16
Clinical Query broad, tw 36 116
Most specific, MeSH 9 9
Most sensitive, MeSH 35 20
Revised Cochrane, MeSH 60 45
Clinical Query narrow, MeSH 23 8
Clinical Query broad, MeSH 35 71

Table 2. Search results for injuries in construction industry material; number of retrieved articles = 31
SEARCH STRATEGY sensitivity % NNR
Original search for CSR 90 136
Most specific, tw 84 21
Most sensitive, tw 84 75
Revised Cochrane, tw 84 73
Most specific, MeSH 32 2