How do I do a systematic review, end to end?
A systematic review runs in six stages: define and register a specific question, search comprehensively and reproducibly, screen against pre-set criteria in duplicate, extract the data you need with the source recorded, appraise each result for bias, and only then synthesise and report. What makes it systematic is that the rules are written down before you see the results — not that any single stage is done in a particular tool.
1. Define the question and register the protocol
Pick the framework that matches the question, then write the eligibility criteria until two people could apply them to the same abstract and agree. Pre-specify the outcomes, the effect measure, the synthesis model, and any subgroups.
Register the protocol before screening starts. Registration is what converts your later analysis choices from "what we found interesting" into "what we said we would do", and it is the single cheapest thing you can do to make the review defensible.
2. Search comprehensively and reproducibly
Search several databases rather than one, because coverage differs and no single source is complete. Build the strategy from the concepts in your question, combining controlled vocabulary with free text.
Record the full strategy per database, with the date run and the number of records returned. Supplement with citation chasing — the references of your included studies, and papers citing them — and with trial registries, which surface studies that were run and never published.
3. Screen against the criteria, in duplicate
Two independent reviewers screen titles and abstracts, then full texts, resolving disagreement by discussion or a third reviewer. Duplicate screening is not ceremony: single screening misses eligible studies at a rate that is easy to measure and uncomfortable to publish.
Record a reason for every full-text exclusion. Those reasons are a required part of the flow diagram, and reconstructing them afterwards is miserable.
4. Extract data with its source
Extract into a form piloted on a few studies first. Capture what you need for synthesis — effect estimates and their precision, or the raw numbers to compute them — plus the study characteristics you will use to judge similarity and to define subgroups.
Record where each number came from: the table, the figure, or the supplement. When two reviewers disagree, or a reader queries a value two years later, the provenance is the only thing that settles it quickly.
5. Appraise each result for bias
Appraise per result, not per study, using a tool matched to the design. This feeds the certainty rating and can justify a sensitivity analysis restricted to low-risk results.
6. Synthesise, rate certainty, and report
Pool only what is conceptually poolable, using the model you pre-specified. Report the pooled estimate with its interval, the between-study variance and a prediction interval, and investigate heterogeneity through the subgroups you named in the protocol.
Rate certainty per outcome, and report against the relevant reporting guideline: the flow diagram with numbers at each stage, the full search strategies, the characteristics of included studies, the excluded-with-reasons list, and a summary of findings table.
If a synthesis is not appropriate, say so and synthesise narratively with structure. A meta-analysis of studies that should not have been combined is more misleading than no meta-analysis, because it attaches a confidence interval to it.
Where this answer stops
This is the shape of the process, not a substitute for the reporting guidance and handbooks that specify it. A review that follows these stages can still be uninformative if the question was not worth asking or the literature cannot answer it.