How to Write a Systematic Review Protocol

A systematic review protocol is a pre-specified plan written before the review is conducted. It's the foundation of rigorous, transparent, reproducible research. Writing the protocol first: - Prevents "outcome switching" (changing your research question after seeing the results) - Forces you to think through your methods before encountering potential biases - Allows peer review of your methods before you invest weeks of work - Creates accountability when registered in databases like PROSPERO Without a protocol, systematic reviews are vulnerable to the same biases they're designed to overcome.

Every systematic review starts with a well-structured clinical question using the PICO framework: - **P**opulation: Who are you studying? (e.g., adults with type 2 diabetes) - **I**ntervention: What treatment or exposure? (e.g., metformin) - **C**omparison: What are you comparing to? (e.g., placebo or other drug) - **O**utcome: What are you measuring? (e.g., HbA1c reduction at 6 months) A well-formed PICO question might be: "In adults with type 2 diabetes (P), does metformin (I) compared to placebo (C) reduce HbA1c by ≥1% at 6 months (O)?" Your PICO determines everything else — your search strategy, inclusion/exclusion criteria, and data extraction form.

Pre-specify exactly which studies you will include and exclude: **Include:** - Study types (e.g., RCTs only, or also cohort studies) - Population characteristics (age range, diagnosis, setting) - Minimum follow-up duration - Outcomes reported - Language restrictions (if any) **Exclude:** - Case reports and editorials - Studies below a minimum sample size - Studies with a high risk of bias (define how you will assess this) - Duplicate publications Be as specific as possible. Vague criteria lead to inconsistent screening decisions.

A comprehensive systematic review searches multiple databases: - **PubMed/MEDLINE**: Essential for biomedical topics - **Embase**: Especially for European clinical trials - **Cochrane Central Register**: Randomized trials - **CINAHL**: Nursing and allied health - **ClinicalTrials.gov**: Unpublished or ongoing trials (reduces publication bias) For each database, develop a search string using: - MeSH terms (controlled vocabulary) AND free-text keywords - Boolean operators (AND, OR, NOT) - Truncation (*) and wildcards Document your exact search string — it must be reproducible. Tools like MetaLens AI can help with initial scoping before you commit to a full search.

Before you start extracting data, design your extraction form. For each included study, you'll typically record: - Study ID, author, year, country - Study design and follow-up - Population characteristics (sample size, age, sex, baseline severity) - Intervention details (dose, duration, comparator) - Outcome data (means, SDs, event rates, effect estimates, CIs, p-values) - Risk of bias assessment Pilot your form on 2-3 studies before full-scale extraction. Two reviewers extracting independently with arbitration reduces errors.

Pre-registering your protocol increases transparency and credibility: - **PROSPERO** (prospero.york.ac.uk): The most widely used registry for systematic reviews - **Open Science Framework** (osf.io): Suitable for any research type - **Cochrane**: If conducting a Cochrane review Registration gives you a datestamped record showing your methods were decided before you saw the data. Most high-impact journals now expect or require registration for systematic reviews. Once registered, any deviations from your protocol must be reported and justified in your paper.

AI tools like MetaLens AI are valuable for the scoping phase — before you write your formal protocol: - Rapidly scan the existing literature to assess whether a systematic review is warranted - Identify the key papers and journals in your area - Understand the current state of evidence and likely effect sizes - Refine your PICO question based on what's actually been studied However, AI tools do not replace a formal systematic review. They work from abstracts, may miss relevant studies, and cannot perform formal risk-of-bias assessments. Use them to inform and accelerate your protocol development, not to replace it.