Skip to content
StudyLens

Our Methodology

A transparent look at how StudyLens extracts, assesses, and contextualizes health research.

Extraction Pipeline

Every study goes through a five-step automated pipeline, with each step idempotent and retryable.

1

Resolve Identifiers

Given a DOI, PMID, or title, we query OpenAlex, Crossref, PubMed E-utilities, and Semantic Scholar to resolve all cross-identifiers and gather maximum metadata.

2

Check Open-Access Status

Unpaywall API determines whether legal full text is available. We never bypass paywalls or use copyrighted content.

3

Fetch Full Text

PMC XML is preferred, followed by OA location XML/plain text, then OA HTML, then DOI landing page HTML as a last resort. Each source is parsed for scholarly metadata.

4

Extract Facts & Quality Flags

Claude AI extracts 28 fact categories (16 core + 12 interrogation) with source citations via structured tool calling. If full text yields zero facts, the pipeline automatically retries with the abstract.

5

Store with Citations

Results are stored with the study record, facts are bulk-inserted with source offsets, and the coverage badge is assigned based on data depth.

Coverage Badges

Each study receives a badge indicating the depth of data available for analysis.

Gold — facts extracted from full text. Silver — facts from abstract or abstract available with registry data. Bronze — metadata only, limited analysis.

Live Example

Full Analysis
Abstract Only
Metadata Only

Quality Checklist

Seven flags give a quick overview of study methodology and reporting quality.

The checklist assesses peer review status, sample size, control group presence, preregistration, study duration, funding disclosure, and conflict of interest declaration. Each flag is data-driven — no subjective judgment.

Live Example

Quality Checklist

Peer Reviewed
Sample Size
Control Group
Preregistered
Duration
Funding Disclosed
Conflicts Declared

Study Interrogation

13 deeper flags across 4 tiers probe for methodological issues that simple checklists miss.

TierFlagsExamples
Critical3Sample size adequacy, p-hacking risk, selective reporting
Important4Effect size, confidence intervals, control group quality, blinding
Moderate3Funding bias, conflict of interest, attrition
Informational3Generalizability, follow-up duration, replication status

Live Example

Study Interrogation

Critical
Important
Moderate
Informational

Caution Meter

A composite score (0-100) aggregates all interrogation flags into a single at-a-glance indicator.

Scores map to four labels: Low Concern (0-15), Some Concerns (16-35), Moderate Concern (36-60), and High Concern (61-100). The meter highlights the top contributing flags.

Live Example

Caution Meter

Low Concern
LowHigh
Follow-up Duration: 6-month follow-up may miss long-term effects.

Caution Meter

Moderate Concern
LowHigh
Multiple Comparisons: 12 secondary endpoints tested without Bonferroni correction.
Industry Funding: Primary funder manufactures the tested intervention.

Confidence Levels

Each extracted fact carries a confidence indicator based on source quality and extraction certainty.

High — extracted from explicit statement in full text. Medium — inferred from abstract or partial information. Low — derived from metadata or indirect sources.

Live Example

High confidence
Medium confidence
Low confidence

Media Discovery

An automated agent finds and analyzes news coverage of each study.

After extraction completes, an LLM agent generates search queries from study metadata, searches Google News RSS, fetches candidate articles, and validates matches with a confidence threshold above 0.7.

Validated articles are classified by stance: positive, negative, cautious, or neutral. A periodic cron re-runs discovery for recent studies (daily for the last 7 days, weekly for the last 30).

For more on our ethical approach to media analysis, see our language standards.

Meta-Topics

Studies are automatically grouped into LLM-managed research areas.

During extraction, the LLM classifies each study into existing topics or proposes new ones. Topics start in a proposed state and are promoted to active once 3 or more studies are associated.

Topic descriptions, hierarchy, and evidence summaries are all LLM-generated. No manual curation is required — the taxonomy evolves as new studies are added.

Standards Alignment

Our quality assessment draws from established reporting standards.

CONSORT Reporting standard for randomized controlled trials
STROBE-nut Reporting standard for nutritional epidemiology
PRISMA Reporting standard for systematic reviews and meta-analyses

Questions about our methodology? Get in touch. For our ethical commitments, see Ethics & Transparency.