SciSpace 指南：如何一步步做系统综述（含真实 PRISMA 示例）

A systematic review is not simply a literature review with more papers.

What makes it systematic is the method: the question is defined, the search is documented, eligibility rules are established, screening decisions are recorded, data are extracted consistently, and the conclusions can be traced back to the included studies.

Drawing on systematic-review sessions by Dr. Ullah and Marija Đukić, this guide focuses on the decisions researchers need to make at each stage. It uses Dr. Ullah’s real review, which narrowed 4,634 retrieved records to 74 included studies, as a worked example.

The systematic review process at a glance

1. Confirm that the systematic review is needed

Before opening multiple databases, establish why another review should exist.

A new systematic review may be justified because:

• The previous review is outdated
• Important new studies have appeared
• Existing reviews cover a different population, context, intervention, or outcome
• Earlier reviews used limited databases or restrictive criteria
• The evidence remains inconsistent
• No review has addressed the specific question
• A methodological or theoretical gap remains unresolved

Start by locating recent reviews on the topic. Compare their questions, publication dates, included studies, databases, limitations, and recommendations for future research.

Marija Djokic uses this approach before beginning a new review: she examines existing systematic reviews and their limitations to determine whether a new contribution is possible.

Write the reason for the review in one sentence

Use this structure: This systematic review is needed because [what the existing reviews do not establish], particularly for [population, context, intervention, outcome, or period].

This sentence does not need to become the final wording in the paper. Its purpose is to test whether the review has a defensible reason to proceed.

Questions to answer before continuing

• What has already been reviewed?
• When was the most recent relevant review published?
• What will this review examine differently?
• Is there enough primary research to review?
• Which methodology and reporting guidance are expected in the field?
• Who needs the findings, and what decision could they inform?

If these questions do not produce a clear contribution, refine the scope before investing in the full review.

2. Define the review question and protocol

The review question determines what will be searched, screened, extracted, and synthesised.

A topic such as “generative AI in higher education” is too broad. A review question needs to identify the specific relationship or evidence being examined.

Depending on the field, a framework such as PICO or PICOC can help organise the question:

Do not force a question into PICO when another framework better fits the review. The framework should clarify the question, not distort it.

Test the question with a pilot

Dr. Ullah recommends testing the proposed question against a small group of approximately eight to ten papers that are already known to be relevant. The purpose is to establish whether the question can be answered and whether the available evidence is likely to produce a useful result.

During the pilot, ask:

• Can the papers provide the information needed?
• Are the concepts defined consistently?
• Is the question too broad?
• Is it so narrow that very little evidence exists?
• Can the same eligibility rules be applied across the papers?
• Which information would need to be extracted from every study?

Revise the question before proceeding when the pilot exposes a problem.

Build a review protocol

The protocol records the decisions made before the results of the review are known.

A protocol does not prevent every change. It makes changes visible.

When the question, criteria, or synthesis method needs to change, document what changed, when it changed, and why.

3. Build a reproducible search strategy

A systematic search should allow another researcher to understand where you searched, what you searched for, and when the search was run.

Begin by dividing the question into concept blocks. For each concept, record:

• Preferred terminology
• Synonyms
• Alternative spellings
• Abbreviations
• Older terminology
• Related terms used in the discipline
• Controlled vocabulary or subject headings, where available

Combine alternatives within a concept using OR. Connect different concepts using AND.

Illustrative search structure

("generative AI" OR ChatGPT OR "large language model*")
AND (student* OR undergraduate* OR postgraduate*)
AND (adoption OR acceptance OR "behavioural intention")

Review each term against known relevant papers. If the search fails to retrieve papers you already know should appear, investigate what terminology those papers use.

Adapt the search to each database

The same string should not automatically be copied into every database. Databases differ in:

• Field codes
• Phrase searching
• Truncation symbols
• Controlled vocabulary
• Search limits
• Indexing
• Document coverage

Marija emphasises that database selection should reflect the field. PubMed may be central to a medical review, while engineering or computing reviews may require databases such as IEEE Xplore or Scopus.

Keep a complete search log

Preserve the exact searches rather than only describing them generally.

“Relevant papers were searched in major databases” is not reproducible. The reader needs to know which databases, which strings, which dates, and which limits were used.

Check the search before screening

Before finalising the search:

1. Confirm that known relevant papers appear.
2. Inspect a sample of irrelevant results.
3. Identify concepts causing excessive noise.
4. Check whether important synonyms are absent.
5. Ask a subject librarian or information specialist to review the strategy where possible.
6. Save the final version used in each database.

A complicated search is not necessarily a good search. Its value lies in retrieving the required evidence transparently.

4. Set inclusion and exclusion criteria before screening

Eligibility criteria convert the review question into rules that can be applied to individual studies.

Define them before examining the complete search results. Otherwise, it becomes easier to change the rules in response to studies you prefer or expect to find.

Inclusion and exclusion criteria template

Avoid criteria such as: 'Include relevant studies.'

Instead write: 'Include empirical studies involving students enrolled in higher education that examine adoption, acceptance, or use of generative AI tools.'

The second version can be applied during screening.

Test the criteria

Apply the draft criteria to a small sample of papers.

Check whether:

• Two reviewers would reach the same decision
• Each criterion can be evaluated from the available information
• Multiple criteria contradict one another
• Important studies are being excluded for an unintended reason
• The rules are too broad to reduce the search results meaningfully

Revise ambiguous criteria before screening the complete dataset.

5. Screen the studies and document the PRISMA flow

Study selection usually involves:

1. Combining records from the selected sources
2. Removing duplicates
3. Screening titles and abstracts
4. Retrieving full texts
5. Screening the full texts
6. Recording exclusion reasons
7. Identifying any studies found through citation searching or other methods
8. Confirming the final included set

The exact order can vary by protocol. Report the sequence that was actually followed.

Worked PRISMA example

Dr. Ullah's manually conducted review produced the following screening pathway:

'This is a real case example, not a recommended target. Another review may retrieve far fewer or far more records.

In Dr. Ullah's example, title screening took place before duplicate removal. Many review teams remove duplicates earlier. What matters is that the protocol and final report state the order used and preserve the counts at every stage

Record full-text exclusion reasons

A final number does not explain why studies were excluded. Use specific reasons such as:

• Wrong population
• Wrong intervention or exposure
• Ineligible study design
• Required outcome not reported
• Conference abstract without sufficient data
• Duplicate or extended version of another study
• Full text unavailable
• Publication outside the specified date or language criteria

Where more than one reason applies, decide how the review will assign the primary exclusion reason and use the approach consistently.

Do not confuse abstract screening with full-text screening

An abstract can establish that a paper appears eligible. It may not contain enough information to confirm the decision.

Marija identifies this problem in her demonstration: a study was treated as included even though its full text was unavailable, which she marked for later checking.

When full text is unavailable:

• Mark the record clearly
• Attempt legitimate retrieval routes
• Do not describe the study as fully assessed
• State how unresolved records were handled
• Include the limitation in the final report where it affects the review

6. Extract data that answer the research questions

Data extraction should not become a collection of whatever information looks interesting while reading.

Define the extraction form before completing the full set of included studies.

Every field should help:

• Describe the evidence base
• Answer a review question
• Assess comparability
• Evaluate methodological quality
• Support the planned synthesis

Data-extraction form

Dr. Ullah's extraction form maps particular data fields directly to individual research questions. This makes it possible to answer the questions from the completed evidence table rather than reconstruct the review from scattered notes later.

Pilot the extraction form

Test the form on several included studies before extracting the complete dataset.

Check whether:

• Two reviewers interpret the fields consistently
• The required information is normally available
• The categories are too broad
• Important details are missing
• The form captures distinctions required for the synthesis

Revise the form before full extraction. Record any later changes.

Separate evidence from interpretation

Use different fields for:

• What the study reports
• What the reviewer concludes about it

For example:

Reported finding: Participants reported greater confidence after the intervention.

(is not the same as)

Reviewer interpretation: The intervention improved performance.

Confidence and performance are different outcomes. Keeping them separate prevents an interpretation from being recorded as if it were a study result.

7. Appraise and synthesise the evidence

The extraction table organises the studies. The synthesis explains what the studies establish together.

First, assess study quality or risk of bias using the method specified in the protocol and appropriate to the included study designs.

Then compare the studies by:

• Population and context
• Definitions
• Interventions or exposures
• Comparators
• Outcomes
• Measurement methods
• Study design
• Sample size
• Methodological strengths and limitations
• Direction and size of findings
• Relevance to the review question

Dr. Ullah's warns against writing the review as a sequence of descriptions: "Study one did this. Study two did this. Study three did this."

The researcher needs to explain the relationships between those studies, including where findings agree, where methods differ, and how the strengths or limitations affect the conclusion.

Illustrative comparison

A weak synthesis would say: 'Study A found high adoption. Study B found reliability concerns. Study C found lower use.'

A stronger synthesis would say: 'Positive intentions did not consistently translate into sustained use. Study A measured students’ stated intention, while Study C examined actual behaviour and found declining use. Study B offers a possible explanation by identifying concerns about reliability, although its qualitative design does not establish that those concerns caused the decline.'

The stronger version:

• Distinguishes intention from behaviour
• Connects the studies
• Preserves methodological differences
• Avoids claiming causality that the evidence does not establish

The synthesis method should match the question and data. Explain why the selected analytical approach is appropriate rather than choosing it only because it is familiar.

8. Report the review transparently and verify every output

A systematic review should allow readers to trace the process from the original question to the final conclusion.

Final review checklist

• The need for the review is explained
• The research questions are stated clearly
• The protocol and later amendments are documented
• Databases and additional search methods are identified
• Exact search strings, dates, limits, and record counts are retained
• Duplicate removal is documented
• Eligibility criteria are reported
• Title, abstract, and full-text screening stages are distinguishable
• Full-text exclusion reasons are available
• Abstract-only and full-text assessments are not treated as equivalent
• Extraction fields are defined
• Quality or risk-of-bias methods are stated
• The synthesis method is justified
• PRISMA counts match the underlying screening records
• Findings can be traced to the included studies
• Limitations and deviations are reported
• Any use of AI is disclosed where required

A polished report can still contain methodological gaps.

In Marija’s demonstration, the generated report included many expected sections but omitted the future-research component she wanted. She also distinguishes procedural tasks that AI can assist with from theoretical development, argumentation, and final interpretation, which remain the researcher’s responsibility.

Where AI can reduce the workload

AI can support procedural parts of a review, but it cannot own the methodology.

The tool can reduce repetitive work. The research team remains accountable for the evidence and conclusions.

Build a review another researcher can follow

A strong systematic review is not defined by the number of papers it contains or how quickly its report was generated.

Its strength comes from the traceability of its decisions: why the review was needed, how the evidence was found, why each study was selected, what information was extracted, and how the conclusions were reached.

Use tools to reduce the repetitive work. Keep the methodological decisions, critical appraisal, synthesis, and final interpretation with the research team.

About the experts

Marija Đukić

Marija Đukić is a Teaching Assistant in the Department of Information Systems at the University of Belgrade’s Faculty of Organizational Sciences. Her guidance strengthens the methodological-rigour sections of this guide, particularly protocol design, database selection, threats to validity and the decisions that still require expert judgment. FON

Her session distinguishes the procedural stages AI can support from theoretical development, argumentation and final interpretation, which remain the researcher’s responsibility. How to Conduct a Systematic Lit…

Dr. Faheem Ullah

Dr. Faheem Ullah is an Assistant Professor at the University of Adelaide, Australia. His manually conducted systematic review provides the central worked example in this guide, showing how a defined question, multi-database search, screening process, data-extraction form and critical synthesis connect in practice.

His review moved from 4,634 retrieved records to 74 included studies, giving researchers a real example of how study selection progresses from database search to final inclusion.

Frequently asked questions

注意转载授权 · 阅读原文