Clinical Literature Review for Evidence-Based Medicine

“Clinical literature review” is an umbrella term; “systematic review” is a specific method with defined requirements. A systematic review answers a pre-defined question using a protocol written and registered before the search, a search reported in enough detail for someone else to run it, independent screening by two reviewers, formal risk-of-bias assessment and a certainty rating. A clinical literature review may be any of six types: systematic, rapid, scoping, umbrella, narrative or a targeted review prepared for a regulatory submission. Each is legitimate under its own name. Problems arise only when the label does not match the method, for example a narrative review presented with a flow diagram, a scoping review used to estimate an effect, or a single-reviewer review described as systematic. What makes a review systematic is not the number of databases but the fact that the protocol was written before the evidence was reviewed.

Follow these steps in order, because the order is part of the method. Frame the question so it can be searched: PICO for effect questions, SPIDER for qualitative questions and PIRD for diagnostic accuracy. Decide which review type the question and your timeline support. Write a protocol to PRISMA-P and register it on PROSPERO before screening. Build the search in each database’s own vocabulary and test it against studies known to belong in the review. Screen with two reviewers independently, recording every full-text exclusion with its reason. Extract into a piloted form, with outcome data extracted in duplicate. Appraise with the instrument suited to the studies. Then synthesise, narratively or by meta-analysis depending on what the studies support, and grade certainty per outcome. Changing the order undermines the review: adjusting inclusion criteria after the results are visible produces a review that no longer supports the claims it makes about its own method.

For a systematic review, registration is expected in practice, and it should take place before screening begins rather than before submission. Registration supports every claim the review makes about its method: that the question was fixed, that the inclusion criteria were not adjusted as the results appeared, and that the reported outcomes are those planned, in the order planned. The registration date is permanently visible in PROSPERO, so a review registered after screening started cannot support those claims, and reviewers check registration dates. PROSPERO accepts systematic, rapid and umbrella reviews but not scoping reviews, which are registered on the Open Science Framework instead. Deviating from a registered protocol is normal and happens in most reviews; what matters is reporting each deviation with its reason.

The instrument depends on the studies being appraised, not on the type of review. Randomised trials: RoB 2, judged per outcome rather than per study, because the same trial can be at low risk for mortality and high risk for a subjective outcome. Non-randomised studies of interventions: ROBINS-I, with the version stated, because ROBINS-I V2 has circulated as a draft and is not yet settled. Exposure studies: ROBINS-E. Diagnostic accuracy: QUADAS-3, which superseded QUADAS-2 in February 2026; a review submitted now using QUADAS-2 will be queried. Systematic reviews, in an umbrella review: AMSTAR 2. Qualitative research: a CASP checklist, with GRADE-CERQual for confidence in the findings. We do not use the Jadad scale. Cochrane advises against summary quality scores, because combining several domains into one number hides which domain was the problem, and domain-based assessment replaced them more than a decade ago.

Databases should be chosen for the question rather than maximised in number. For a review of randomised evidence in a biomedical field, MEDLINE, Embase and the Cochrane Library cover most of the literature, with trial registries for completed but unpublished trials. Scopus or Web of Science add coverage outside the biomedical core and allow forward citation searching. CINAHL is important for nursing and allied health, PsycINFO for mental health, and regional databases such as LILACS, African Journals Online and CNKI are more important than many English-language reviews acknowledge. Adding further databases does not improve a review if the same poorly translated strategy is run in each. Two well-built searches retrieve more than five copied ones, because the vocabularies differ: MeSH in MEDLINE and Emtree in Embase are different thesauri, and a MEDLINE strategy pasted into Embase returns plausible results from the wrong set. Report which databases you searched and, equally, which you did not search and why.

One person can produce an excellent review, but not one that meets the usual requirements for a systematic review, because independent dual screening is one of them. Single-reviewer screening misses eligible studies at a well-documented rate, and the missed studies are not random: they cluster among records that are ambiguous from the abstract. There are two appropriate options. Conduct the work as a rapid review and state in the methods which systematic review steps were abbreviated and why, which is a recognised and widely published approach. Alternatively, bring in a second reviewer for the screening stage only, which is the least costly part of the review to outsource and the stage where a second reviewer most affects the result. Single screening should never be described in a way that leads readers to assume otherwise.

First check that it is the field and not the strategy, which is why testing against known studies before running matters. If the search is sound and the field really is thin, that is a finding rather than a failure, and it is often a more useful one than the review you were expecting. A well-documented empty or near-empty review tells a guideline panel that no recommendation can be made, tells a funder exactly where the gap is, and tells a trialist what the next study should be. Empty reviews are published and are cited. What does not work is widening the inclusion criteria until something appears, because the criteria were registered, the change is visible, and a review whose scope expanded after the results were seen cannot claim what a systematic review claims. Where we find this we will say so and set out the honest options: report the gap, reframe the question deliberately and re-register, or stop.

No, and treating them as the same is expensive in both directions. A literature review inside an EU MDR clinical evaluation, an IVDR performance evaluation, an HTA submission or a regulatory dossier is a controlled document within a technical file. It is judged by an auditor or an assessor against the submission’s own template and expectations, it carries a version history rather than a publication date, and it has to be re-runnable to a stated date because the obligation recurs — a clinical evaluation is updated, not finished. It also frequently has to appraise evidence about a specific device or product rather than answer a general effect question, which changes the search entirely. A systematic review written to PRISMA is not wrong in this setting, and PRISMA is genuinely the clearest available structure for reporting the search, but it answers a different question and takes roughly twice as long. Conversely, a regulatory literature review submitted to a journal will be asked why there is no protocol registration.

Published estimates for academic teams run to a year or more, though much of that is elapsed time rather than work — people screening in the evenings between other commitments. As concentrated work at the scale most clinical reviews run to, budget something like three weeks of reviewer time across all stages, with screening the largest single block and full-text retrieval the step that always takes longer than anyone plans for. The variables that actually move the number are the size of the search, whether screening is dual, how many studies survive to extraction, and whether authors have to be contacted for missing data. The chooser at the top of this page gives an estimate for your parameters, and our free screening workload estimator breaks the same figure down stage by stage, with every rate as an editable field, so you can adjust our estimates. Reviews most often overrun when a search returns four thousand records to find forty studies; time spent with an experienced information specialist at the start saves more time than faster screening.