Title and Abstract Screening You Can Defend in a Methods Section
Screening is not “the AI ranked papers.” It is pre-specified eligibility, an inspectable first pass on titles and abstracts, and a human who can reverse every recommendation.
A literature review becomes a method the first time a paper is kept or dropped. Everything before that is search. Everything after that is synthesis of a set you already chose. If the choice cannot be inspected, the rest of the manuscript is decorating a black box.
LitSynth’s screening stage is a title and abstract first pass. That is a defensible method when you report it as such. It is not dual independent screening, and it is not a completed PRISMA flow diagram.
The citable tutorial lives in the screening methods page. This essay is the longer argument: what to write, what not to invent, and why “maybe” is a valid screening outcome.
The method, in one paragraph
You can paste this into a Methods section:
Title and abstract screening was conducted against eligibility criteria specified before ranking. LitSynth generated include, exclude, or uncertain recommendations with criterion-level reasons. A human reviewer confirmed or overrode each decision. Dual independent screening and full-text eligibility assessment were not completed inside the software.
If you later did full text or a second reviewer outside the tool, add those sentences. Do not delete the first-pass boundary to sound more systematic.
Eligibility criteria have to be able to fail a paper
A criterion that cannot exclude anything is a slogan. Write population, intervention or exposure, comparator if the question needs one, outcomes, study designs, years, and language limits as text a second person could apply to the same abstract.
Do this before you look at relevance scores. Changing the criteria after seeing which papers would die is a protocol amendment. Record the version. If you changed mid-stream, say so.
A high relevance score is not inclusion. An abstract that “sounds on topic” is not inclusion. The only inclusion rule is: the record meets the criteria you wrote, or a human explicitly overrode the recommendation and wrote why.
Treat the model as a recommendation, not a vote
LitSynth can judge criteria as yes / no / maybe and attach a reason. That is useful because later you can ask “why was this dropped?” It is not useful if you collapse maybe into exclude to make the counts tidy.
Keep three buckets until a person resolves them:
- Include — the abstract appears to meet the criteria.
- Exclude — at least one criterion fails, and the reason is stored.
- Uncertain — the abstract does not contain enough information, or the criterion is borderline.
Uncertain is not indecision. It is the honest title/abstract outcome when the methods reside in full text. Moving those records to exclude because you are in a hurry invents a completed screen you did not do.
An exclusion you cannot inspect is a deletion, not a method
The screening record worth keeping is: decision, reason, criterion that failed, and whether a human overrode the suggestion. Export it. If you only keep the included pile, you cannot reconstruct the flow, you cannot audit the AI, and you cannot answer a reviewer who asks why a landmark trial is missing.
Two further bookkeeping rules:
- Do not mix screening exclusions with extraction failures. A paper dropped at title/abstract is not the same object as a card that could not be extracted.
- Do not silently drop borderline records. If the human disagreed with the model, that override is part of the method.
Title/abstract counts are not a PRISMA diagram
PRISMA wants identification, title/abstract screened, full-text assessed, full-text excluded with reasons, and studies included in the synthesis. LitSynth can support the title/abstract slice. It does not, by itself, complete the rest.
Report what you have:
- records retrieved and after deduplication, if you have those numbers;
- title/abstract screened;
- included / excluded / still uncertain at this stage.
Do not invent “full-text excluded, n = …” from titles. Do not call the first pass a systematic review because the software used the word screening.
The review workflow tutorial exists so that screening, extraction, and appraisal stay in order. Extraction of records you have not screened is busywork. Appraisal of titles is theatre.
What LitSynth does not do here
Be explicit in the same paragraph as the tool:
- It does not implement dual independent screening.
- It does not complete full-text eligibility inside the product.
- It does not turn a first pass into a finished PRISMA counts table.
- A relevance rank is not an inclusion decision.
If your protocol requires two reviewers, run the first pass in LitSynth, export the record, and complete the second reviewer (and full text) outside. Cite both steps.
How to read a screening table without fooling yourself
- Open the criteria, not the scores. If a criterion is vague, the whole column is vague.
- Sample the excludes. A method that only inspects includes will ratify whatever the model already liked.
- Keep maybes visible. If the uncertain column is empty on a hard clinical question, someone probably forced a binary.
- Write the override. “I included it anyway” without a reason is not documentation.
When screening is done, extraction can start. The next essay is how to extract an evidence table you can inspect. Residual research questions come even later, in the Research Gaps Matrix—and only from evidence you actually included.