A systematic review is less about being clever and more about being disciplined. The method is a fixed sequence of steps, each one documented so that someone else could follow your trail and land in the same place. Skip a step or fudge one, and the whole thing loses the credibility that separates it from an ordinary literature review.
This guide walks through the process end to end, in the order you actually do it. The stages overlap in practice and you will loop back more than once, but the sequence below is the backbone. Expect the full project to run somewhere between six months and well over a year, depending on the size of the literature and your team.
Everything downstream depends on this. A systematic review answers one specific question, not a broad topic. "Exercise and depression" is a topic. "In adults with major depressive disorder, does aerobic exercise compared with usual care reduce depressive symptoms?" is a question.
The PICO framework is the standard tool for building one: Population, Intervention, Comparison, Outcome. Different fields use variants, such as PECO for exposure questions or PICo for qualitative ones, but the goal is the same. You want a question narrow enough that you can decide, for any given study, whether it belongs.
If you cannot yet phrase your question this tightly, the field may be too broad or too new for a systematic review, and a scoping review might be the better starting point.
Before you search anything, write down your plan. The protocol states your question, your databases, your search terms, your inclusion and exclusion criteria, your data extraction plan, your appraisal tool, and your planned analysis. It exists to stop you from unconsciously bending the rules once you see the results.
Register it publicly, usually on PROSPERO for health-related reviews. Registration time-stamps your plan, so no one can later accuse you of changing your criteria to get a tidier answer. Many journals now expect a registered protocol, and following PRISMA-P helps you write a complete one. This step feels like bureaucracy. It is actually your main defense against your own bias.
Now you search for every study that could be relevant. This is deliberately broad. You would rather sift through some irrelevant hits than miss an eligible study.
Search several databases, not one. Which ones depend on your field, but health reviews commonly cover MEDLINE (via PubMed or Ovid), Embase, and the Cochrane Library, with others like CINAHL or PsycINFO as the topic demands. Build your search from your PICO concepts, combining terms with Boolean operators and controlled vocabulary such as MeSH where it applies. Use truncation and synonyms so you catch variant spellings and phrasings.
Record everything: which databases, which terms, which dates, how many hits from each. The search has to be reproducible, so document it in enough detail that a reader could paste it in and get your numbers. Then extend beyond databases by checking reference lists of key papers and, where relevant, grey literature and trial registries.
A multi-database search returns the same studies more than once. Before screening, pull everything into a reference manager or review platform and remove duplicates. This sounds trivial and is not; deduplication errors ripple through your PRISMA counts and make the flow diagram inconsistent later.
Keep a clean record of how many results you started with and how many remained after deduplication. Those numbers feed directly into your reporting.
Screening is where you narrow the pile down to the studies that actually meet your criteria, and it happens in two passes.
First, screen titles and abstracts against your inclusion criteria to discard the clearly irrelevant. Then retrieve the full text of everything that survives and screen those against the same criteria in detail. At the full-text stage you record a reason for every exclusion, because you will report those reasons.
The standard is to have two reviewers screen independently at both stages, then resolve disagreements by discussion or a third reviewer. This is not busywork. Single-reviewer screening misses eligible studies and lets bias creep in. Measure your agreement with a statistic like Cohen's kappa if your field expects it.
For every included study, pull out the information you planned to collect: study characteristics, participant details, methods, and the outcome data you need for synthesis. Design the extraction form during the protocol stage and pilot it on a few studies first, because you will always discover a field you forgot.
Have two people extract independently where feasible, and reconcile differences. If a study reports data in a form you cannot use directly, such as a median instead of a mean, you may need to convert it or contact the authors. Keep track of what you converted and how.
A systematic review does not just count studies; it judges how much to trust each one. Assess every included study with a formal tool matched to its design. Randomized trials commonly use Cochrane's RoB 2. Non-randomized studies use ROBINS-I. Observational designs might use the Newcastle-Ottawa Scale, and diagnostic accuracy studies use QUADAS-2.
This appraisal shapes how you interpret everything that follows. A result built on high-risk studies deserves less confidence than the same result from well-conducted ones, and your synthesis should say so. Many teams visualize the results as a traffic-light plot so readers can see the pattern at a glance.
Now you bring the evidence together, and how you do that depends on the studies.
If they measured comparable outcomes in comparable populations, you can pool them statistically in a meta-analysis, producing an overall effect size, a forest plot, and a heterogeneity statistic such as I-squared to show how much the studies disagree. If the studies are too different to combine, you synthesize them narratively instead, describing patterns and contradictions across the evidence in a structured way.
Choosing not to pool is a legitimate and often correct decision. A misleading forest plot is worse than an honest narrative synthesis. Let the data decide, not the desire for a single number.
Beyond appraising individual studies, judge how much confidence to place in the body of evidence for each outcome. The GRADE approach is the common framework, rating certainty as high, moderate, low, or very low based on factors like risk of bias, inconsistency, imprecision, and publication bias. The result often goes into a summary of findings table. This step tells readers not just what the evidence says, but how far they should trust it.
Finally, write it up against a reporting standard, almost always PRISMA. The PRISMA checklist makes sure you report every element a reader needs to judge and reproduce your work, and the PRISMA flow diagram accounts for every record from your initial search down to the final included studies.
Structure the manuscript so the methods are detailed enough to repeat, the results present your synthesis and appraisal clearly, and the discussion states your findings, their certainty, and the review's limitations honestly. Then choose a journal that publishes systematic reviews in your field and submit.
| Step | What you do | Key output |
|---|---|---|
| 1 | Frame a focused question | A PICO question |
| 2 | Write and register a protocol | Registered protocol (e.g. PROSPERO) |
| 3 | Build and run the search | Documented, reproducible search |
| 4 | Deduplicate and organize | Clean record set with counts |
| 5 | Screen in two stages | Included studies, exclusion reasons |
| 6 | Extract the data | Completed extraction forms |
| 7 | Appraise risk of bias | Risk-of-bias assessments |
| 8 | Synthesize | Narrative synthesis or meta-analysis |
| 9 | Assess certainty | GRADE ratings, summary of findings |
| 10 | Report and publish | PRISMA-compliant manuscript |
A few mistakes account for most of the trouble. Starting to search before the protocol is written, which invites unconscious rule-bending. Searching only one database, which guarantees missed studies. Screening with a single reviewer, which lets bias and error slip through. Pooling studies that are too different, which produces a confident but meaningless number. And underestimating the timeline, then cutting corners near a deadline. Each of these is avoidable, and each is the difference between a review that holds up and one that gets picked apart.
The steps are simple to describe and heavy to run. You are juggling thousands of records, two-reviewer screening logs, extraction forms, appraisal tables, and a PRISMA diagram whose numbers all have to reconcile. Managing that across spreadsheets and email is where errors get in.
Verflux runs the whole pipeline in one place: search imports and deduplication, independent screening with conflict resolution, data extraction, risk-of-bias appraisal, and meta-analysis, with the PRISMA flow diagram built automatically from your real screening decisions. The method still has to be yours, done properly at each step. The platform just keeps the numbers consistent so the mechanics do not become the hard part.
How long does a systematic review take? Usually six months to over a year. The biggest drivers are the number of records to screen, the size of your team, and whether you run a meta-analysis.
Can I conduct a systematic review alone? It is difficult, mainly because the standard expects two independent reviewers for screening and extraction. Solo reviews exist, but you should be transparent about the trade-offs, and most guidance recommends a team.
Do I have to register the protocol? It is strongly recommended and required by many journals. Registration protects you against accusations of changing your plan after seeing the results, so there is little reason to skip it.
How many databases should I search? More than one. The exact set depends on your field, but a single database will miss eligible studies. Health reviews often search at least MEDLINE, Embase, and the Cochrane Library.
Does every systematic review need a meta-analysis? No. Pool studies only when they are similar enough to combine. Otherwise a structured narrative synthesis is the correct approach.
Conducting a systematic review means following a documented sequence: frame the question, register the plan, search comprehensively, screen and extract in duplicate, appraise the evidence, synthesize it, rate its certainty, and report it against PRISMA. The discipline is the point. Each documented step is what lets a reader trust your answer and, if they wish, retrace it.
Get the protocol right, respect the two-reviewer standard, and let the data decide whether to pool. Do that, and you will produce a review that stands up to scrutiny.
Planning your first systematic review? Verflux walks you through screening, appraisal, and meta-analysis in one workflow, with the PRISMA diagram handled for you.
All analytical features included in the free trial. No credit card, no installation, no R or Python.
Create free account