How to write a literature review

1. Frame the review question

A review starts with a question, not with a search. "Machine learning in medicine" is a topic; "which machine learning methods are used for early sepsis detection and how well are they validated on external cohorts" is a review question. The more specific the question, the easier the screening and the more useful the result.

Fix the boundaries early: period, types of work (empirical, review, theoretical), languages, disciplines. These boundaries later become the search method section.

2. Find sources beyond keywords

Keywords find what is named with familiar terms. To avoid missing adjacent work, add two techniques: search by the meaning of the question and follow citations. From a few strong papers go backward to their references and forward to the papers that cite them.

Record where and how you searched: databases, queries, dates. For a systematic review this is mandatory; for a narrative review it protects you from "why did you ignore…".

• Keywords and their synonyms, including terminology from neighbouring fields.
• Semantic search by the question wording to catch work with different vocabulary.
• Backward and forward citations from 3–5 anchor papers.
• Recent publications from the last year or two that older reviews do not cover.

3. Screen and read with one template

Screening happens in two passes: title and abstract, then full text. Write the inclusion and exclusion criteria before reading, otherwise they drift.

Read with a single template and fill a table: the paper's question, method, sample or data, key result, limitations, what the authors suggest next. That table is the skeleton of the review.

4. Structure by theme, not by author

The most common mistake is a review of the form "Smith showed… Jones found… Lee proposed…". That is an annotated list, not a review. Group sources by theme, method, point of disagreement or the chronology of an idea, and compare the works within each group.

A good test: every paragraph has its own claim and the citations support or challenge it. If a paragraph starts with a surname, there is probably no claim.

5. Synthesise: agreement, contradictions, gaps

Synthesis answers three questions. Where do researchers agree and how strong is the evidence? Where do results diverge and what explains it: different samples, methods, definitions? What is unstudied or weakly studied?

The third question is what makes the review useful for your own work: the gaps justify its relevance.

6. Write the introduction and conclusion last

The introduction describes the question, the boundaries and the search method. The conclusion sums up the synthesis and states what your study will do. Both are easier to write once the body exists.

7. Format references and check each one

Use the style your institution or journal requires: APA, Vancouver, GOST and others. Check that each reference really supports the claim it is attached to rather than being merely on-topic. Exporting to BibTeX or RIS removes manual formatting.

Common mistakes: a review without a question, retelling by author, no recent sources, no critique, references that are "roughly on topic".