How to do a Literature Review | Step by Step | PhD Thesis Writing

A literature review becomes manageable when it’s built like a pipeline: start with thesis-level headings, break each heading into focused sub-questions, then search, screen, summarize, and synthesize until each paragraph has clear claims backed by clustered evidence. The core idea is to avoid drowning in papers by narrowing the scope to one subtopic at a time—so the review grows in organized layers rather than as a chaotic pile of reading notes.

The process begins by drafting the chapter structure for the thesis, including likely “main headings” and related-work sections. Instead of treating these as final, the approach treats them as working placeholders. From there, each main topic is converted into a set of subtopics framed as questions. For an injury prediction and prevention section, example sub-questions include how frequently runners get injured (and how that varies by distance or runner type), what risk factors drive those injuries, which prevention techniques already have support, and what advances exist in computer science that could enable prediction using machine learning.

Once a subtopic is chosen—say, “how frequent are running-related injuries?”—the next step is keyword brainstorming. The goal is to generate search terms that capture both the population and the outcome, such as “running,” “marathon,” “runner/marathoner,” and injury incidence concepts like “incidence,” “frequency,” or “injury incidence.” With keywords in hand, the workflow shifts to database searching (the transcript mentions PubMed as an example, but the method is database-agnostic). Search results can be overwhelming, so screening happens in two passes: first the title, then the abstract. Papers that don’t match the subtopic are rejected early, and selected papers are stored in a structured folder system (e.g., a folder for “injury prevention,” with subfolders for each subtopic).

After selecting papers, the researcher creates a lightweight summary document populated with brief, subtopic-specific findings—typically only a couple of sentences per paper—so patterns can emerge without writing full paragraphs too soon. Then comes citation tracing: using tools like Google Scholar (and similar platforms) to follow who cited the chosen papers and which references they cite. This step helps expand the set of relevant literature, but it also requires restraint; citation tracing can explode into thousands of documents, so systematic reviews should often be treated as sufficient unless a specific referenced study is directly needed.

Reading follows the same scope logic. For “frequency of injuries,” the emphasis is on results and what the studies found, not on replicating the study methods. For computer science subtopics—where machine learning approaches may be central—the methodology sections become more important because the thesis may depend on understanding how those techniques work.

Finally, synthesis turns summaries into paragraph-ready claims. The transcript illustrates aggregating multiple papers into a handful of main points, such as: injury incidence is high across runner distances; ultra marathoners show high rates of injuries requiring medical care; time-loss injuries follow a U-shaped relationship with running distance; novice runners experience the highest injury rates; and injury incidence relates to dropping out of running. Each paragraph is then shaped with a topic sentence and a concluding sentence that highlights the remaining gap—often setting up the next section (e.g., moving from incidence to risk factors and then to prevention). The overall payoff is a repeatable method for building a literature review one subtopic at a time, with evidence organized enough to write confidently.