About Blindspot

1. Data Sources

Blindspot queries the following databases in real time for every search. No results are pre-indexed; all counts and review lists are fetched live and cached for up to 7 days.

Deduplication

Reviews retrieved from multiple sources are deduplicated by DOI, PubMed ID, and normalised title before display. The primary study count appliesID-based deduplication — Blindspot samples IDs from each source and computes the actual overlap fraction rather than applying a fixed estimate. This approach adapts to the specific query: queries with high PubMed/OpenAlex overlap apply a stronger deduplication factor than queries with low overlap. Typical inter-database overlap: PubMed/OpenAlex 50–70%, PubMed/Europe PMC 40–60%.

Living Systematic Reviews

Blindspot identifies “living systematic reviews” (LSRs) — continuously updated reviews that incorporate new evidence as it emerges. If one or more LSRs are found on your topic, an informational banner is displayed indicating that the research area may already be under active review. This is particularly relevant for rapidly-evolving clinical topics where an LSR program may already be addressing the gaps you identified.

OpenAlex Coverage Limitation

A 2025 PMC study found that OpenAlex has 96% recall for a systematic review benchmark set, but 3–4 records per search may be missing due to closed-access abstract removal (March 2025). Blindspot's per-source study count breakdown (visible on any result page) allows researchers to compare database counts directly.

2. Feasibility Scoring

Feasibility scoring is entirely data-driven — no AI is used. The score is calculated from the blended unique primary study count returned by the database queries. Thresholds are aligned with Cochrane Handbook recommendations:

What counts as a primary study?

The primary study count excludes systematic reviews, meta-analyses, reviews, and protocols. Only original research articles (randomised controlled trials, observational studies, cohort studies, etc.) are counted. This is enforced via PubMed publication type filters and OpenAlex work-type exclusions.

Query broadening for count accuracy

PICO-structured inputs are used to generate a targeted AND-Boolean query for existing review searches. The same query in plain form (without Boolean operators) is used for primary study counting, to avoid under-counting studies that use synonymous terminology. A synonym expansion library maps common concept pairs (e.g., "adolescents" ↔ "youth", "myocardial infarction" ↔ "heart attack") to improve recall.

3. Where AI Is — and Is Not — Used

AI is used for: Gap analysis

After the database searches complete, Blindspot passes the titles and abstracts of up to 20 existing reviews to Google Gemini 2.0 Flash for gap analysis. Gemini identifies potential research gaps across six dimensions: population, methodology, outcome, geographic, temporal, and theoretical. It also suggests specific review titles and a study design recommendation.

Gap analysis confidence is indicated in the interface — High Confidence when 20 reviews were analyzed, Moderate Confidence for 10–19, Low Confidence for 5–9, and Very Low Confidence for fewer than 5. When fewer than 3 primary studies exist (Insufficient feasibility), AI gap analysis is blocked entirely because there is insufficient evidence for the AI to analyze meaningfully.

AI is used for: Alternative topic suggestions

When a topic returns Low or Insufficient feasibility, Blindspot uses the OpenAlex Topics taxonomy to find sibling research areas in the same academic subfield. Each candidate is verified against PubMed's real study counts before being shown. Only topics with 6 or more primary studies (Moderate feasibility threshold) are suggested.

AI is NOT used for: Feasibility scoring

The feasibility score (High / Moderate / Low / Insufficient) is derived entirely from database query counts. No AI model estimates or adjusts the primary study count or the resulting score.

AI is NOT used for: Study counts or review retrieval

All study counts and existing review listings are retrieved directly from PubMed, OpenAlex, Europe PMC, and Semantic Scholar APIs. AI does not fabricate, estimate, or supplement these results.

AI is NOT used for: Registry checks

PROSPERO and OSF registry checks are direct API queries. Blindspot does not use AI to infer whether a review is registered.

4. PRISMA Flow Diagram Estimates

Blindspot generates an estimated PRISMA 2020 flow diagram showing how many records would be identified, screened, and included in a systematic review on the searched topic. These are estimates, not actual review numbers.

Screening ratios are calibrated against published systematic review datasets from a 2021 meta-research study (Bannach-Brown et al.) covering 100 Cochrane reviews and 200 non-Cochrane systematic reviews. Estimates are stratified by corpus size into five tiers (XS / S / M / L / XL) to reflect the different screening yields observed across different evidence base sizes.

Included counts are presented as a range (95% confidence interval) rather than a point estimate, reflecting uncertainty in screening ratios across review types. The displayed range is conservative: the interval spans approximately ±35% of the central estimate for small corpora and ±20% for large corpora.

Search telemetry (logged since March 2026) will allow retrospective calibration of these ratios against real Blindspot searches once sufficient data accumulates.

5. Known Limitations

6. Human Oversight Requirement

Blindspot outputs are advisory. All AI-generated gap analysis, study design recommendations, and alternative topic suggestions must be verified by a qualified researcher before being used in a systematic review protocol, grant application, or publication.

Feasibility scores reflect a snapshot of the indexed literature at the time of the search. A High feasibility score does not guarantee that a systematic review is appropriate or that no review exists outside the indexed databases. Researchers should independently search grey literature, conference proceedings, and discipline-specific databases before finalising a review topic.

PROSPERO and OSF registry checks identify registered protocols but do not guarantee that an unregistered or embargoed review is not in progress. Researchers should contact field experts and check specialist registries relevant to their discipline.

7. RAISE 3 Alignment

Blindspot is designed to align with Cochrane RAISE 3 (June 2025) — guidance on selecting and using AI evidence synthesis tools responsibly. The table below summarises alignment with the RAISE 3 disclosure checklist:

For questions about methodology, data sources, or institutional use, please contact us via the feedback link in the application footer.