Strata Academy

Literature Screening: Title/Abstract & Full-Text Methods

Dual review, exclusion reasons, PRISMA flow, Covidence/Rayyan workflow, and limits of AI screening

1. Why screening is a separate methodological step

Comprehensive search strategies intentionally cast a wide net – MEDLINE, Embase, CENTRAL, trial registries, grey literature – and often return thousands of references after deduplication. Screening applies inclusion and exclusion criteria consistently to decide which reports proceed to full-text assessment and eventual synthesis.

Biased screening distorts the evidence base as much as a weak search. If one reviewer unconsciously favours positive abstracts from high-impact journals, or excludes non-English titles without justification, the review's conclusions may reflect selection bias rather than the true literature.

Screening is where eligibility criteria meet reality. A beautifully written PICO in the protocol means little if full-text exclusions are not recorded with PRISMA reason codes.

When appraising someone else's review, scrutinise the flow diagram before reading the discussion. Large unexplained drops between screened and included records are an immediate quality concern.

2. Two-stage screening process

Stage 1 is title and abstract screening against inclusion/exclusion criteria. Reviewers mark records as include, exclude, or unsure. Unsure records proceed to full text rather than being silently dropped.

Stage 2 is full-text retrieval and assessment. Not every abstract-stage include will meet criteria once methods are visible. Excluded full-text papers require a documented reason per PRISMA – wrong design, wrong population, wrong intervention, wrong outcomes, or insufficient data.

Track 'full text not retrieved' separately from 'excluded after full text'. Non-availability is a limitation that may introduce publication bias if many studies are missing.

Pilot screening on 50–100 records calibrates team understanding before the full set. Revise the protocol if criteria prove ambiguous in practice.

• Common exclusion reasons – wrong population, intervention, comparator, outcome, study design
• Conference abstracts may proceed to full text if they might contain usable data
• Duplicate reports of the same trial should be linked before extraction
• Resolve disagreements before extraction begins

3. Dual independent review

Cochrane methods expect independent screening by two trained reviewers, with a third resolver for conflicts. This applies to both title/abstract and full-text stages in rigorous reviews.

Single screening is a methodological limitation that should be acknowledged and may warrant downgrading confidence in the review's conclusions. Some rapid reviews justify single screening for feasibility – students should recognise this trade-off.

Blinding reviewers to authors, institutions, and journal names reduces prestige bias where feasible. Covidence and similar tools support blinding settings.

Calculate chance-corrected agreement (e.g. Cohen's kappa) on the pilot sample if your course requires reliability metrics. Low agreement signals unclear criteria, not 'bad reviewers'.

4. Screening software

Covidence, Rayyan, DistillerSR, and Abstrackr support duplicate screening, exclusion labels, conflict resolution, and PRISMA flow exports. Choose one tool early and document the workflow in your protocol.

Reference managers (EndNote, Zotero) handle deduplication before import. Deduplication rules should be reported – matching on DOI, PMID, and author-year reduces false duplicates.

Export screening decisions in a format that preserves reviewer identity and timestamps. This audit trail supports supervision, publication, and PRISMA compliance checks.

For very large searches, some teams split records by alphabet or database only after randomisation – not by predicted outcome – to avoid systematic bias in workload allocation.

5. AI-assisted screening – benefits and limits

Active learning classifiers can prioritise likely-included abstracts in reviews with tens of thousands of records. Reviewers screen high-probability records first, retraining the model as labels accumulate.

AI does not remove the need for human inclusion decisions in rigorous reviews. False negatives – relevant studies never shown to a human – are the primary risk. Conservative workflows use AI for ordering, not silent exclusion.

Training data bias, opaque models, and validation on non-representative samples can mis-rank records. Report tool name, version, training approach, sensitivity targets, and how many records received full human review.

Generic LLM chat is not a substitute for structured screening logs and PRISMA counts. Chat cannot produce reconciled dual-reviewer flow diagrams auditable by an examiner.

• Suitable – prioritisation within dual human screening at scale
• Unsuitable – sole reviewer replacing humans for final inclusion in high-stakes reviews
• Always report – tool name, validation metrics, human oversight model
• Re-validate when changing clinical topic or search strategy

6. Linking screening to PRISMA

PRISMA 2020 expects counts at every stage: records identified, deduplicated, screened, excluded at title/abstract, sought for retrieval, not retrieved, assessed at full text, excluded with reasons, and included in synthesis.

Every excluded record at full-text stage needs a reason category for the flow diagram. 'Other' should be minimised – if overused, your criteria were probably unclear.

Numbers must reconcile across abstract, flow diagram, supplementary appendices, and prose methods. Examiner and reviewer suspicion rises quickly when totals do not add up.

When appraising a review, rebuild the flow mentally: if thousands were screened but only two exclusion reasons appear, request the full exclusion log.

7. Common screening mistakes

Starting full-text retrieval before completing dual title/abstract screening – duplicates workload and blurs audit trails.

Excluding conference abstracts without attempting to find full publications linked to the same trial.

Changing eligibility criteria mid-screening without protocol amendment.

Using single screening while claiming Cochrane-equivalent methods.

Failing to document translator support for non-English records when the search was language-restricted.

Interactive version (quizzes, walkthroughs) loads when JavaScript is enabled.