How To Do A Systematic Review And Meta-Analysis In Medicine | Fast, Clear Steps

Medical decisions lean on trusted summaries of the best trials. A well-run systematic review and meta-analysis brings the evidence together with clarity. This guide sets out a practical route that you can follow from idea to publication, with tips that match standard methods and reporting checklists.

Doing A Systematic Review And Meta-Analysis In Medicine: Step-By-Step

Here is the big picture. Build a team, write a protocol, search broadly, screen and extract in duplicate, appraise bias, run transparent statistics, rate certainty, and report with a clear flow diagram.

Map The Question And Outcomes

Keep the question narrow enough to allow a useful synthesis yet broad enough to matter in practice. Spell out PICO and add time points and settings. Decide the primary outcome that drives your sample size and main figures. List secondary outcomes that fill out the picture. Define study designs that fit the question, such as randomized trials for treatment effects or cohort studies for harms.

Write And Register A Protocol

The protocol is your contract with readers and with your later self. It guards against changing methods after looking at results. Most teams publish the protocol in a registry that tracks health reviews. The PROSPERO portal explains what to include and how to submit a record. Include aims, PICO, eligibility, search plan, screening process, extraction items, risk of bias tool, effect measures, subgroups, and plans for any updates.

Design A Reproducible Search

Work with an information specialist. Combine subject headings and free text with tested synonyms, including drug and brand names. Search multiple databases plus trial registries and gray sources to reduce non-publication bias. Log platform, date, years spanned, filters, and full strings. Save histories and exports so others can repeat your steps.

Core Sources To Include

Most medical topics need MEDLINE (via PubMed or another host) plus Embase. Add CENTRAL for trials, CINAHL for nursing or allied topics, and Web of Science for citation chasing. Trial registries such as ClinicalTrials.gov and WHO ICTRP help you find recent or unpublished trials. Hand-search main journals and conference proceedings when indexing is patchy.

Manage Records And Screening

Export all records, merge, and de-duplicate before screening. Write simple, testable criteria for titles and abstracts. Train reviewers on a sample and measure agreement. Use two reviewers for each record. Disagreements go to a third person or to a consensus chat. Record every reason for full-text exclusion with one clear label per study. You will need those labels for the flow diagram and for transparency.

Extract Data Consistently

Design a form that captures study design, setting, sample size, baseline balance, intervention details, comparator details, and outcomes with time points and analysis type. Note adjustments, per-protocol or intention-to-treat, and any protocol deviations. Pilot the form on three to five studies and refine it before full use. Extract in pairs, checking each field, and contact authors for missing values when needed.

Assess Risk Of Bias

Judge methods, not results. For randomized trials, use RoB 2 with its domains on randomization, deviations, missing data, measurement, and selection of the reported result. For non-randomized studies of interventions, use ROBINS-I and align confounders with your question. Document a record for each judgment and keep a record of signaling answers. Create traffic-light plots or summary figures to help readers see patterns.

How To Conduct A Systematic Review And Meta Analysis In Medicine: Practical Walkthrough

Before pooling, check that studies answer the same question. Compare PICO, outcome definitions, and time points. Standardize units if needed, and avoid double counting when a trial has more than two arms. Pre-specify the effect measure for each outcome and the model you plan to use.

Pick Effect Measures That Fit The Data

For binary outcomes, use a single choice across studies (risk ratio or odds ratio). For time-to-event data, use hazard ratios. For continuous outcomes on one scale, use mean difference; when scales differ, use standardized mean difference. Keep direction aligned so higher always means the same thing.

Model Choice, Heterogeneity, And Small-Study Effects

Random-effects models suit most medical topics. Report how you estimated between-study variance and show a confidence interval for the pooled effect. Describe heterogeneity with τ² and I² and add a forest plot. Probe small-study effects with a funnel plot and a simple regression test when study counts allow.

Handle Missing And Complex Data

Prefer adjusted estimates that match the question. Convert medians to means only when samples are large and distributions look near normal. Adjust cluster trials for design effects. For crossover trials, use paired analyses or first-period data if carryover is likely. List any imputation rules and test them in sensitivity checks.

Plan Subgroups And Sensitivity Checks

Pick a short set of subgroups rooted in science, such as dose, risk level, setting, or study quality. Label them exploratory unless named as confirmatory in the protocol. Sensitivity checks can drop high risk studies, switch models, swap effect measures, or run leave-one-out analyses.

Rate Certainty With GRADE

Readers need to know how much trust to place in each pooled estimate. Rate certainty across five domains: risk of bias, inconsistency, indirectness, imprecision, and publication bias. Start high for randomized trials and lower the rating when the domains point that way; start lower for non-randomized designs. Present a summary of findings table that shows the pooled effect, baseline risk, absolute effect, and the certainty rating.

Plan a one-page summary of findings table for decision makers. List the primary outcomes first, then the ones that shape safety or convenience. For each outcome, report the study count, total sample, the pooled relative effect with its 95% CI, the baseline risk you chose, and the absolute effect over a meaningful time frame. Add the certainty rating and short footnotes that state the domain and reason for any rating change, such as high risk of bias or wide intervals. Link the table in the abstract for quick scanning by busy readers worldwide. Put this table near the main text so readers see the take-home message before diving into plots and appendices.

Report With Clarity

Core Reporting Items

Follow the PRISMA 2020 checklist from title to tables. Describe the search so another team can repeat it. Add a flow diagram with counts and reasons for each full-text exclusion. Provide tables for study features, bias, and results. Share the extraction form, raw data, and code in a public link.

Statistical Notes That Save Time

State one primary outcome in advance and stick with it. Pre-define the window for each time point. When pooling proportions, consider logit or arcsine transformations to stabilize variance and avoid false precision from tiny cells. Scale continuous outcomes to units that matter at the bedside. Add absolute effects so busy readers can grasp impact at a glance.

Software And Reproducible Workflows

RevMan can handle many standard tasks. R packages such as metafor, meta, and gemtc add flexible models and clear code. Keep a script that builds your figures and tables from raw data. Store the script and the data with version control so your team can audit each change.

Keep Bias Low At Every Stage

Pairs of reviewers, clear forms, and blinded first passes reduce errors. Invite a content expert and a methods lead to review tricky calls. Document each protocol deviation and why it was needed. Independent extraction for the primary outcome pays off in better trust.

Lean On Authoritative Methods

The online Cochrane Handbook gives step-by-step guidance on searches, bias tools, models, and interpretation. Use it as your day-to-day reference while you draft methods and plan analyses.

From Draft To Submission

Write the methods first, then results, then the plain language take-home points. Keep figures readable in black and white. Put all search strings in an appendix. Add a data link in the manuscript and on the registry record. Ask a colleague outside the team to rerun your main analyses from your files.

Plan For Harms, Subgroups, And Equity

Safety outcomes deserve the same care as benefits. Define adverse event terms in advance and stick with one scheme. Note whether studies used active monitoring or passive reports. Track serious events and withdrawals due to events as separate outcomes. Present both relative and absolute risks.

Subgroup claims can mislead when study counts are small. Keep lists short and tie them to prior science. Check for differences with a formal test instead of scanning separate rows.

Record who was studied. When data allow, show results by sex, age band, and setting. Even when no pooling is possible, a table that shows coverage by group helps readers judge fit.

When Meta-Analysis Is Unwise

Not every data set should be pooled. If populations, comparators, and outcomes do not match, or if reporting hides core inputs, a meta-analysis may create false confidence. In that case, use a structured narrative. Group studies by design and outcome window, describe directions and sizes, and explain what blocks pooling. The SWiM principles for synthesis without meta-analysis offer a tidy path.

Network And Indirect Comparisons

Some topics compare many active treatments with few head-to-head trials. A network meta-analysis can help by linking direct and indirect evidence. Before you fit a model, check that study populations and settings can be seen as part of one network, and that effect modifiers are shared across links. Report the geometry of the network, pick a random-effects model, and show both league tables and rank plots with clear caveats.

Update And Maintain Your Review

A good review ages as new trials appear. State how you will track updates, such as monthly alerts in core databases and trial registries. Keep a change log in your repository so readers can see what changed and when. When new evidence shifts a primary outcome or the certainty rating, post an updated dataset and a short note in the registry record.

Short Checklist Before You Submit

• Title names the question, design, and population.
• Protocol registered and linked from the paper.
• Search strings and dates provided in full.
• Screening done in pairs with reasons for exclusion recorded.
• Data extraction tested, then done in pairs.
• Bias judged with the right tool for each design.
• Effect measures and models stated in advance.
• Forest plots, funnel plots, and a summary of findings table included.
• Certainty rated with GRADE and footnotes that cite reasons.
• Data and code shared in a stable public link.