How to Conduct a Systematic Literature Review

A systematic literature review is a documented process for answering a focused question from the full body of eligible evidence. The essential qualities are not volume or automation. They are a protocol decided in advance, a reproducible search, explicit inclusion decisions, verified extraction, and a synthesis that preserves uncertainty.

This guide gives you an end-to-end workflow. It is suitable for researchers planning a formal review and for teams that want more discipline in a smaller evidence scan. For intervention reviews, use it alongside the Cochrane Handbook and the reporting requirements in PRISMA 2020.

1. Define the question before searching

Start with the decision your review should inform. A broad topic such as “digital interventions for anxiety” is not yet a review question. Specify the population, intervention or exposure, comparator, outcomes, and eligible study designs.

For intervention questions, the PICO framework provides a useful structure:

• Population: who or what is being studied?
• Intervention: what treatment, exposure, or policy is evaluated?
• Comparator: what is the alternative?
• Outcome: what measurable result matters?

Write a one-sentence objective and test it against edge cases. Would you include adolescents? Observational studies? Conference abstracts? A study with six weeks of follow-up? Ambiguity here becomes disagreement during screening.

2. Write a protocol and eligibility criteria

The protocol is the review’s decision log before results can influence those decisions. At minimum, record:

1. the review question and rationale;
2. inclusion and exclusion criteria;
3. information sources and planned search dates;
4. screening and conflict-resolution procedures;
5. data fields to extract;
6. risk-of-bias tools;
7. outcomes and effect measures;
8. the planned synthesis method;
9. sensitivity or subgroup analyses, if justified.

If you change the protocol later, record what changed, when, and why. A justified amendment is more credible than silently adapting the method after seeing the results.

3. Build a reproducible search strategy

Translate each concept in the question into controlled vocabulary and free-text synonyms. A biomedical query might combine a population block, an intervention block, and a study-design block with AND, while synonyms inside each block use OR.

For example:

("major depressive disorder" OR depression)
AND (saffron OR "Crocus sativus")
AND (randomized OR randomised OR placebo)

Do not optimize the query only for precision. A systematic search usually accepts more irrelevant records to reduce the chance of missing eligible studies. Pilot the query against a small set of known relevant papers: if those papers do not appear, revise the vocabulary or filters.

Record the exact query for every database, the platform used, all limits, and the final search date. The PRISMA-S extension provides a reporting checklist specifically for literature searches.

SinaPilot Discovery can search PubMed and Europe PMC from a plain-language topic or an advanced query. Use it to assemble and rank a candidate set, then preserve your final database queries separately for reproducibility.

4. Deduplicate and screen in two stages

Merge search results and remove exact and probable duplicates before screening. Keep a stable record identifier so that decisions remain traceable even if citation metadata changes.

Screen in two stages:

1. Title and abstract screening: remove records that clearly fail the criteria.
2. Full-text screening: assess the remaining reports against every eligibility rule.

Use a short list of standardized exclusion reasons at full text, such as wrong population, wrong intervention, ineligible design, or unavailable outcome. Avoid a vague label such as “not relevant.” If two reviewers screen independently, define how disagreements will be resolved before screening starts.

The count at each stage feeds the PRISMA flow diagram. Keep excluded full-text records and their reasons rather than reconstructing them at manuscript time.

5. Extract data into a structured evidence table

Create the extraction form before reviewing every paper. The columns should map to your question and planned synthesis, not to whatever details happen to be easy to copy.

Useful fields include:

Pilot the form on two or three different studies. This exposes ambiguous fields before the whole dataset has been extracted. For each numerical result, capture the estimate, its uncertainty, the analysis population, and the time point together. A p-value without those elements is rarely enough.

A structured research paper summary workflow helps with first-pass orientation, but any AI-assisted extraction must be checked against the paper, tables, supplements, and registry record.

6. Assess risk of bias separately from reporting quality

A poorly reported study is difficult to assess, but incomplete reporting is not identical to high risk of bias. Use a tool appropriate to the study design and judge the domains it specifies rather than assigning a vague overall “quality score.” The risk-of-bias assessment guide explains how the unit of assessment and tool change across trials, non-randomized intervention studies, diagnostic studies, and reviews.

Typical concerns include allocation, deviations from intended interventions, missing outcome data, outcome measurement, and selective reporting. Preserve the evidence supporting each judgment. If the paper does not provide enough information, mark the domain as unclear or use the tool’s prescribed category rather than guessing.

The same principle applies during critical appraisal of a clinical trial: distinguish what the study found, how trustworthy that finding is, and whether it applies to your question.

7. Compare studies before combining them

Before any synthesis, compare the studies’ PICO elements, design, follow-up, outcome definitions, and risk of bias. Two studies that both mention “response” may use different thresholds and time points. Pooling them without resolving that difference creates a precise answer to an unclear question.

Build an evidence matrix and group studies only when the grouping has a defensible clinical and methodological rationale. The practical process in how to compare research papers shows how to normalize terms and expose contradictions.

If meta-analysis is appropriate, define the effect measure and model, examine heterogeneity, and run planned sensitivity analyses. If it is not appropriate, describe the actual synthesis method. The Cochrane Handbook warns that the label “narrative synthesis” alone is not enough and discourages vote counting based only on statistical significance.

8. Report the review so another team can audit it

PRISMA 2020 asks authors to report the rationale, objectives, eligibility criteria, information sources, full search strategies, selection process, extraction process, risk-of-bias methods, synthesis methods, and results. Use the PRISMA 2020 checklist while writing, not only before submission.

Your final report should make four paths visible:

• how records entered and left the review;
• how each included study contributed data;
• how judgments and transformations were made;
• how the evidence supports the conclusion and its uncertainty.

Common systematic literature review mistakes

• Searching before fixing eligibility criteria. Results then shape the question.
• Using one convenient database. Coverage varies by field and source type.
• Changing outcome definitions during extraction. Decide and document rules first.
• Treating every report as an independent study. One trial can produce several papers.
• Equating “not significant” with “no effect.” Examine estimates and confidence intervals.
• Letting AI make unverified eligibility decisions. Automation should leave an audit trail and a human decision.
• Writing study-by-study paragraphs without synthesis. Organize evidence around comparisons and outcomes.

A reusable review checklist

Before you conclude, confirm that you can answer yes to each question:

• Is the review question explicit and operational?
• Were methods established before results were known?
• Can another researcher reproduce every database search?
• Does every excluded full text have a reason?
• Can every extracted value be traced to its source?
• Was risk of bias assessed with a design-appropriate method?
• Are study groupings and synthesis choices justified?
• Does the conclusion reflect certainty, inconsistency, and missing evidence?

The goal is not a frictionless review. It is a review whose decisions remain understandable months later—and defensible to someone who did not make them.

Related guides

• Turn a broad topic into searchable concepts with the PICO framework for research questions.
• Standardize extraction with an accurate research paper summary.
• Organize a body of evidence using a research paper comparison matrix.
• Challenge individual studies with a structured peer-review checklist.
• Turn unresolved evidence into a defensible research gap statement.