Systematic reviews are a cornerstone of evidence-based care, either by themselves or through their incorporation into practice guidelines, performance measures or other evidence-based practice. To be useful, however, systematic reviews need to be up-to-date.

The science of determining when systematic reviews need updating has been developing for the past decade. Prior to 2001, no method or criterion existed to determine whether evidence-based products remained valid or whether the evidence underlying them had been superseded by newer work. Since then, several groups have begun developing methods to determine signals for updating reviews [1–5]. Most methods involve some form of limited literature searches and the use of expert opinion, although some methods use statistical methods and are applicable only to meta-analytic results [6, 7]. Two of these methods have been formally compared and found to produce similar results [2]. To date, however, no method has been assessed for predictive validity, meaning there is no way of determining whether the presence or absence of signals does in fact predict whether the review is out-of-date. In addition to the more easily assessed situation of a false-positive (that is, a signal that detects that a review is out-of-date, but the subsequent update does not result in any important changes in the conclusions), such a study requires being able to assess for false-negatives, which requires updating reviews for which no signals are detected. In 2008, we were asked to determine which of 11 systematic reviews sponsored by the Agency for Healthcare Research and Quality (AHRQ) Comparative Effectiveness Review (CER) program might be in need of updating. We took advantage of a natural experiment to assess the predictive validity of our method for assessing for signals for updating.

In this study, we assessed the predictive validity of signals for updating CERs detected in 2009 that have since been updated. We start with a description of the original process used to detect signals [3] and then describe how we assessed the validity of the signals. This original process subsequently evolved to the process described by Ahmadzai et al.[8]; the two are nearly identical.

The great majority (83%) of conclusions for each CER and across CERs had good concordance. However, the CER on gastroesophageal reflux disease (GERD) had four “out-of-date” conclusions with only fair concordance, and one conclusion we had assessed as “still valid” was shown to be out-of-date.

The published 2009 updating assessment judged that the conclusion regarding endoscopic treatment for GERD “should be deleted”, meaning that it was out-of-date, because the endoscopic procedures had been withdrawn from the market. However, one of the three endoscopic procedures reviewed in the original report continued to be used, new endoscopic procedures were introduced and one of the two withdrawn procedures was later reintroduced. The update report noted this changing landscape, and we deemed the concordance with the 2009 prediction as only fair. A more appropriate surveillance assessment would have been that the conclusion needed updating because the endoscopic procedures were evolving over time.

Another conclusion in the original GERD report—that surgery and medical therapy were similarly effective—was rated as “still valid” during the surveillance process but had poor concordance with the update review, which concluded that surgery was favored over medical therapy. One of the studies providing new evidence in support of this conclusion was published in 2009, after completion of the surveillance signal search.

This assessment of the predictive validity of a method to assess a CER for signals for updating yielded generally favorable results. For the vast majority of individual conclusions, concordance between the 2009 predictions and the subsequent updated conclusions was judged to be good. The one instance of poor concordance had new evidence published after the surveillance signals had been assessed, and in this instance involved a CER already judged to be of high priority for updating based on signals of other out-of-date conclusions.

Our present study has three primary limitations. The first is sample size. We were able to assess only nine CERs. However, this number included CERs assessed as being of high, medium or low priority, thus allowing us to assess the possibility of false-negatives (that is, CERs assessed as low priority for updating that nevertheless were fully updated). The likelihood of assessing such false-negatives again is small, as it requires that low-priority CERs be subjected to the gold standard of a full update. Our findings that neither of the CERs judged to be a low priority had any substantive changes in conclusions will reinforce the decision to invest scarce resources in researching other topics rather than commisioning updates of low-priority CERs.

A second limitation is the matching of original conclusions to updated conclusions. In some updated reports, the authors themselves matched the conclusions. In most cases, however, this was not done, and, in some circumstances, determining the appropriate match to the original conclusion was challenging. Additional file 2 lists each original conclusion and its matching updated conclusion so that readers may judge this agreement for themselves.

The third principal limitation of this study is that the 2013 assessment of the 2009 predictions could not be made in a blinded fashion. Our Evidence-based Practice Center (EPC) did both assessments, and, even if some other group had done the 2013 assessment, we could not have enforced blinding, because the 2009 assessments are in the public domain. We tried to guard against bias by having explicit reasons for each judgment and presenting these reasons for readers themselves to judge. Our reasoning should be transparent.

With the limitation of small sample size in mind, we offer the following preliminary conclusions about the surveillance signal method. (1) Low-priority CERs are unlikely to have any substantive changes in conclusions. (2) Conclusions judged likely to be “still valid” almost certainly are still valid. (3) Conclusions judged to be “out-of-date” almost certainly are out-of-date. (4) Safety concerns and the appearance of new classes of therapies and more efficacious treatments are the best targets for high-priority updates. (5) The classification of individual conclusions as possibly or probably out-of-date owing to new evidence may be slightly too sensitive as a signal; in a number of such instances, the update report’s conclusion did not change, because the new evidence identified in the signal search was either rejected or insufficient to change the original conclusion.

In sum, our assessment provides some support for the predictive validity of this method of assessing CERs for signals of the need for updating. Future research is likely to be confined to assessing updates of systematic reviews judged to be a medium or high priority for updating. Further assessment of the factors leading to changes in individual conclusions may help refine the criteria for distinguishing between high- and medium-priority update topics. However, investing extra time and effort to distinguish “possibly” from “probably” out-of-date conclusions or to further refine the global assessment to distinguish medium- from high-priority update topics may begin to make the surveillance process resemble the actual update, which is not the goal of surveillance. In this application, the surveillance process worked very well—nearly perfectly, in fact (κ ≥ 0.8 is considered nearly perfect agreement). No low-priority CER was judged, as having had a substantive change to a conclusion in the update, whereas 3 of 4 high priority CERs did have substantive changes to the conclusions. The results suggest that it is very unlikely that new, practice-changing evidence exists concerning a systematic review judged to be a low priority for updating and supports a policy of delaying an update of a systematic review until new evidence is sufficient to warrant assigning it at least a medium priority.

The assessment method described herein represents part of the basis for the surveillance method used to assess AHRQ systematic reviews as described by Ahmadzai et al. [8]. That program was designed to assess each AHRQ systematic review every 6 months and to take 3 months to complete. One important result is that no systematic review was judged to be a high priority for updating at the first 6-month assessment, meaning that it is probably more cost-effective to assess systematic reviews no more frequently than yearly. Additional work on making surveillance more cost-effective is warranted.

In our present study, we found evidence supporting the predictive validity of a method for assessing AHRQ systematic reviews regarding their need for updating. One advantage of this method relative to other proposed methods is that it is equally useful for meta-analytic reviews and narrative reviews. It may be applicable to systematic reviews produced by other organizations.