How to Conduct a Systematic Literature Review (SLR) and Synthesize Prior Work Like a Senior Researcher

Fig. 1. The six steps of a systematic literature review[cite: 5].

Step 1: Define the Question and Write a Protocol

Everything downstream depends on a sharp question. 'What is known about machine learning in maintenance?' is far too broad. Frame the question with a structure such as PICO(C): population, intervention, comparison, outcome, and context [3]. First, check that an SLR is the right type of review[cite: 5].

Write two to four research questions, such as: RQ1, which architectures are used? RQ2, how are they validated? RQ3, what gaps remain[cite: 5]?

Write a protocol before you search: databases, search strings, criteria, extraction fields, quality questions, and how disagreements are resolved[cite: 5].

Date-stamp the protocol in a repository or on OSF. It proves you did not change the criteria after seeing the results[cite: 5].

Step 2: Search Systematically

• Use at least two or three major databases, such as Scopus, Web of Science, and IEEE Xplore or ScienceDirect[cite: 5].
• Do not rely on Google Scholar alone. Its results are hard to reproduce and cannot be exported cleanly[cite: 5, 4].
• Build the search string from concept blocks. Join synonyms with OR inside a block, and join blocks with AND (Fig. 2)[cite: 5].
• Pilot the string. If it misses five to ten papers you already know are relevant, revise it[cite: 5].
• Record the database, date, string, filters, and number of hits for every search[cite: 5].
• Add backward and forward snowballing to catch papers the databases missed[cite: 5, 5].

Fig. 2. A search string built from concept blocks: OR inside a block, AND between blocks[cite: 5].

Step 3: Screen Studies Transparently

Write your inclusion and exclusion criteria before screening, and apply them the same way to every paper[cite: 5].

• Remove duplicates first, using the DOI or a cleaned title (code below)[cite: 5].
• Screen in two rounds: titles and abstracts, then full texts. Record the reason for every full-text exclusion[cite: 5].
• Tools such as Rayyan [6] and Zotero make this much easier[cite: 5].
• Have two reviewers screen at least a sample independently. Measure agreement with Cohen's kappa [7] and settle disagreements by discussion[cite: 5].

Report the counts in a PRISMA 2020 flow diagram [2] (Fig. 3)[cite: 5, 2].

Fig. 3. A PRISMA-style flow diagram with illustrative numbers[cite: 5].

Step 4: Extract Data and Judge Quality

• Build the extraction sheet before reading in depth. Typical columns: citation, year, problem, method, data, metrics, results, and limitations[cite: 5].
• Pilot the sheet on five papers, then fix it. Changing columns later means re-reading everything[cite: 5].
• Copy numbers exactly as reported, with units, conditions, and page numbers[cite: 5].
• Judge quality with four to six yes/no questions, such as: Are the aims clear? Is the data described? Are the baselines fair? Are limitations stated[cite: 5]?
• Do not hide weak studies. Say how they affect your confidence in the findings[cite: 5].

Step 5: Synthesize Like a Senior Researcher

This is where reviews succeed or fail. A summary says what each paper did. A synthesis says what the papers mean together (Fig. 4)[cite: 5].

Fig. 4. Summary versus synthesis: the second paragraph interprets the pattern (illustrative)[cite: 5].

• Build a concept matrix with papers as rows and themes as columns [8]. The column totals show you what the field actually studies (Fig. 5)[cite: 5].
• Code the papers systematically. Thematic analysis gives a clear routine: read, code, group, and name the themes [9][cite: 5].
• Turn themes into a taxonomy that organizes the methods and shows how they relate (Fig. 6)[cite: 5].
• Compare results on common metrics only when conditions are comparable. When they are not, say so[cite: 5].
• Look for contradictions, and offer reasons: different data, settings, or definitions[cite: 5].
• For large sets, keyword and co-citation maps from VOSviewer can reveal clusters you did not expect [10][cite: 5].

Fig. 5. A concept matrix: papers as rows, themes as columns, totals at the bottom (illustrative)[cite: 5].

Fig. 6. A taxonomy built from the themes found in the reviewed papers[cite: 5].

Find the Gap

Cross two dimensions, such as method and application, and count the studies in each cell (Fig. 7). Thin cells are candidate gaps. Check that a gap matters and is not just empty. A real gap is one a reviewer would agree is worth filling[cite: 5].

Fig. 7. A gap map: methods against applications, with thin rows highlighted (illustrative)[cite: 5].

Step 6: Write It Up and Report It

• Structure: Introduction, Method (databases, strings, criteria, flow diagram), Results (descriptive statistics, then themes), Discussion, Threats to validity, and Conclusion[cite: 5].
• Write each theme as an argument. Open with the finding, then support it with the studies[cite: 5].
• Use tables for details and prose for insight. Never walk through papers one by one[cite: 5].
• Be honest about limits: database coverage, language, publication bias, and subjective coding[cite: 5].
• Tie future directions to the gaps you found[cite: 5].
• Share search strings and the extraction sheet as supplementary files so others can reproduce your work[cite: 5].

Common Mistakes

• Searching only one database, or only Google Scholar[cite: 5].
• Changing criteria halfway without recording it[cite: 5].
• Listing papers instead of synthesizing them[cite: 5].
• Skipping quality appraisal[cite: 5].
• Submitting with a stale search. Re-run it shortly before you submit[cite: 5].
• Ending with vague gaps[cite: 5].

Final Checklist

• Research questions and protocol are written and dated[cite: 5].
• Search strings, databases, dates, and hit counts are recorded[cite: 5].
• Criteria are fixed, and screening is documented in a PRISMA flow diagram[cite: 5].
• Extraction and quality sheets are complete[cite: 5].
• Synthesis shows themes, a taxonomy, and clear gaps[cite: 5].
• Limitations and supplementary files are ready[cite: 5].

A strong review is not the one with the most papers. It is the one a reader can trust, repeat, and build on. Follow the method, show your work, and let the synthesis carry your argument[cite: 5].