Inclusion of methodological filters in searches for diagnostic test accuracy studies misses relevant studies

Abstract

Objective: To compare the performance of MEDLINE searches using index test(s) and target condition (subject searches) with the same searches combined with methodological filters for test accuracy studies.

Study design and setting: We derived a reference set of 506 test accuracy studies indexed on MEDLINE from seven systematic reviews that conducted extensive searches. We compared the performance of "subject" with "filtered" searches (same searches combined with each of 22 filters). Outcome measures were number of reference set records missed, sensitivity, number needed to read (NNR), and precision (Number of reference set studies identified for every 100 records screened).

Results: Subject searches missed 47 of the 506 reference studies; filtered searches missed an additional 21 to 241 studies. Sensitivity was 91% for subject searches and ranged from 43% to 87% for filtered searches. The NNR was 56 (precision 2%) for subject searches and ranged from 7 to 51 (precision 2-15%) for filtered searches.

Conclusions: Filtered searches miss additional studies compared with searches based on index test and target condition. None of the existing filters provided reductions in the NNR for acceptable sensitivity; currently available methodological filters should not be used to identify studies for inclusion in test accuracy reviews.