Evidentum Pro — Systematic Review Workbench

Evidentum Pro

search · screen · appraise

Esc to skip

⚙ Settings

🔒 Privacy & data flow

Everything runs in this browser — there is no Evidentum server. Tap for exactly what leaves your browser, and when.

Your projects, screening decisions, notes and API keys are stored locally in this browser (localStorage / IndexedDB) and are never uploaded by the app. Data leaves your browser only when you act:

Searching sends your query to each database you tick — PubMed/NCBI, Europe PMC, Cochrane, Crossref, OpenAlex, Semantic Scholar, Scopus/Elsevier, ClinicalTrials.gov — directly to that provider’s API.
AI appraisal sends the paper’s title/abstract (and fetched full text) plus the prompt to whichever AI provider’s key you entered (OpenAI, Anthropic, Google, xAI). It runs only when you start it, and is billed to your key.
Semantic Scholar may retry through a public CORS proxy if the direct call is blocked — then your search terms pass through the proxy operator (see the Semantic Scholar section below). No other database search uses a proxy.
Full-text PDF retrieval requests a paper’s PDF directly from its host. Only if that is blocked and you opt in (Full-text PDF retrieval section) does it retry via a public proxy, which then sees the full PDF URL; a signed / credential-bearing link is never proxied. Off by default.
API keys are saved in this browser (localStorage by default; browser extensions or scripts running on this page could read them). Untick “Remember keys on this device” in the AI keys section to keep them for the current tab only — cleared when you close it. Backups exclude keys unless you opt in.

AI output is an advisory draft for a human reviewer to verify — never a final eligibility or risk-of-bias judgement.

🤖 AI Agent — API Keys

Not set model: gpt-5.6

Remember keys on this device. When on, keys stay saved in this browser. Turn off to keep them for this tab only — cleared when you close it.

OpenAI

Responses API. Pick a model above, or choose Custom for any other ID. platform.openai.com →

Anthropic

Messages API. Pick a model above, or choose Custom for any other ID. console.anthropic.com →

Google Gemini

generativelanguage API — large context & low cost, good for high-volume abstract screening. Free key from Google AI Studio →

xAI Grok

OpenAI-compatible API. Key from xAI Console →

Experimental features — OpenAI "Sign in with ChatGPT" (not functional)

OpenAI Account (Sign in with ChatGPT)

Experimental / not yet functional: this flow only establishes identity (openid/profile/email) — the resulting token is not accepted by OpenAI's API and Evidentum does not use it for analysis. All AI analysis runs on the API keys above. OpenAI's third-party sign-in is a gated program — apply via their developer form →

📚 LibKey — Institutional Library Access

One-click full-text access via your institution. Find your ID at libkey.io

📄 Full-text PDF retrieval

Allow a public-proxy fallback when a full-text PDF can’t be fetched directly (off by default)

🔒 Privacy note: When Evidentum reads a paper’s full text for appraisal it requests the PDF directly from its host. If your browser blocks that (CORS) and you tick this box, it retries through a public CORS proxy (allorigins / codetabs / thingproxy) — which means the complete PDF URL is sent to the proxy operator. With the box unticked (the default) a blocked PDF is simply skipped and reported, with no third party in the path. To protect your access, a signed or credential-bearing PDF link (one carrying an access token, signature or institutional-resolver key) is never routed through a proxy even when this is on.

👓 Scopus API Credentials Optional

Free for academic use. Register at dev.elsevier.com · InstToken needed for off-campus access · Enable the Scopus chip on the search tab once keys are entered.

🔬 NCBI / PubMed API Key Optional

Raises PubMed's rate limit from 3 to 10 requests/sec — recommended for large searches (e.g. 500/DB). Free from your NCBI account → Settings → API Key Management. Searches still work without it, just slower.

🧠 Semantic Scholar API Key Optional

Limit Semantic Scholar to biomedical fields (Medicine, Biology)

A free key moves Semantic Scholar off the shared anonymous rate-limit pool (fewer “rate limited” errors) — request one at semanticscholar.org/product/api. The biomedical-fields limit raises precision but can reduce recall (S2's field tags are imperfect), so leave it off for an exhaustive search and on to cut off-topic noise.

Allow the public-proxy fallback for Semantic Scholar (off by default)

🔒 Privacy note: Semantic Scholar usually works with a direct request. Rarely, a browser blocks it (CORS); only then, and only if you tick the box above, Evidentum retries through a public CORS proxy (allorigins / codetabs / thingproxy) — and the full request URL, including your search terms, passes through the proxy operator. With the box unticked (the default) a blocked Semantic Scholar request simply fails and is reported, with no third party in the path. Every other database always talks directly to its API; the desktop build never uses a proxy.

🔑 Saved credentials

All the keys above are stored in this browser only. Wipe every one of them — AI provider keys plus Scopus, NCBI and Semantic Scholar — in a single step. Your projects and screening data are not affected.

🌍 Europe PMC Free · no key

Expand Europe PMC queries with MeSH / UMLS synonyms

Europe PMC's synonym=true auto-adds controlled-vocabulary synonyms (e.g. “heart attack” also matches “myocardial infarction”) — a strong recall boost but a large precision hit (result counts can jump >10×). Leave off for a precise search; turn on to maximise sensitivity. Deep results now page through all matches (cursorMark) instead of stopping at the first 1,000.

📏 Phrase Proximity Advanced

Matches multi-word phrases when the words appear within N words of each other, instead of exactly adjacent — wider recall than an exact phrase, far tighter than a loose AND. Applied to the PubMed ([tiab:~N]), Europe PMC and Semantic Scholar ("…"~N) queries; Scopus / OpenAlex / ClinicalTrials keep exact phrases (no native proximity operator).

💾 Backup & Restore

Your projects, screened papers and screening decisions are stored in this browser only — they're lost if you clear browsing data or switch device. Export a full backup regularly (and before any clean-up), then restore it anywhere. API keys are never included.

↑ Restore backup ⇆ Import reviewer packet

💾 Assessment Memory

AI appraisals are saved locally (by DOI/PMID) so a paper is never re-assessed — or re-billed — across searches. Export a portable pack to back it up, move it to another device, or share it.

…

↑ Import pack

📚 References, Licences & Copyright

Appraisal tools & reporting guidelines — citations & licences

Evidentum Pro is a free, non-commercial research tool. The reporting guidelines and risk-of-bias / quality-appraisal instruments below are reproduced for use within the app with attribution to their authors. Items under a NonCommercial-NoDerivatives licence are reproduced verbatim (signalling questions unaltered) and used only because this tool is non-commercial. Always consult the original source for full guidance.

Using papers responsibly — copyright & the PDF uploader

The articles you search, link to, or upload remain the copyright of their publishers and authors. Evidentum grants you no rights over them — it only helps you find, read and appraise material you are already entitled to access. You are responsible for using each paper within its licence.

✓ Open Access — fine to analyse with AI

When a free / OA full text is found (the green ◆ OA full text badge on a result), AI agent analysis may appraise the whole paper. OA licences (e.g. CC‑BY) explicitly permit reuse and text-and-data mining.

⚠ Subscription / paywalled — read it yourself, don't auto-feed it

An institutional subscription is a personal-use licence: it lets you read the article, via the 📚 Library Access (LibKey) link. Evidentum will not automatically send paywalled text to an AI provider, and you should not bulk-upload subscription PDFs into an automated pipeline — systematic downloading breaches most publisher agreements and can suspend your whole institution's access.

📋 The PDF uploader — what's OK to upload

Only upload a PDF you have lawfully obtained and may use for your own research — your own manuscript, an Open Access article, or a paper you accessed through your institution's subscription for personal study. An uploaded PDF is processed in your browser; its text is sent only to your chosen AI provider for that single appraisal. Evidentum has no server and never stores or forwards it anywhere else.

🤖 Where the text goes

Any full text or abstract you appraise is sent directly to the AI provider you selected (OpenAI, Anthropic, Google Gemini or xAI Grok) under their API and data-handling terms. Review your provider's data-use and retention policy before sending third-party copyrighted content.

This is practical guidance, not legal advice. When in doubt about a specific article or publisher, check with your institution's library or copyright office.

Define your question (PICO)

Enter your research question in PICO terms. The assistant drafts concept blocks and search terms; you edit, restructure, and approve every part. Live result counts come from PubMed’s server — never from the AI.

P — Population / Problem

I — Intervention / Exposure

C — Comparison (optional)

O — Outcome

Study design focus (tick any): RCTs / controlled trials Systematic reviews / meta-analyses Observational studies Diagnostic accuracy

Intervention vs comparator: Either — OR (recommended) Both — AND Also search the outcome concept (Cochrane: usually off)

From date To date

Assembled PubMed strategy & live count

Numbered search sets (PRESS / PRISMA-S — for your methods section)

Paywalled databases — copy & paste (Ovid Embase · Cochrane CENTRAL)

These interfaces have no public search API, so paste the translated strategy into their own search page and run it there; record the retrieved counts in your PRISMA flow alongside the databases searched here. ⚠️ Ovid: subject headings are translated mechanically from MeSH — verify each exp …/ line against Emtree (Embase’s own thesaurus) and remap where the vocabularies differ. CENTRAL indexes MeSH natively so headings carry over as-is; no RCT filter line is added because CENTRAL is already a register of controlled trials.

Ovid · Embase (Advanced Search)

Cochrane Library · CENTRAL (Search Manager)

ℹ️ This PubMed strategy is translated to each database’s native syntax when you send it to the workbench. MeSH headings only search as MeSH in PubMed (and, partially, Europe PMC) — Scopus, Semantic Scholar, OpenAlex and ClinicalTrials.gov search them as free-text keywords, so every concept needs free-text synonyms to travel. Inverted headings (e.g. “Diabetes Mellitus, Type 2”) are auto-rewritten to natural word order for those engines.

The count is the true number of PubMed records this strategy retrieves (NCBI History server). Aim for a manageable, comprehensive set — the coach above suggests adjustments.

🎯 Known-item recall check (optional but recommended)

Paste PMIDs or DOIs of papers you already know belong in this review. If the strategy misses any, it isn’t sensitive enough — widen the terms and re-check. The date filter is ignored here (concept terms + study-design only).

🔑 To unlock full use of Evidentum, an API key from OpenAI, Anthropic, Google Gemini or xAI Grok is needed — please enter it in the panel.

Keyword triage signals

— a reading-order aid, applied to each title & abstract after retrieval

Each signal scans the title and abstract for your keywords. Must contain = flagged ✓ if any keyword is found. Must NOT contain = flagged ✓ if no keyword is found (e.g. "animal, rat, mouse" to down-rank animal studies). More signals detected ranks a paper higher for triage. Keywords match at the start of a word (like a database * truncation), so "randomis" catches "randomised" and "randomisation" but "male" won't match "female". These are simple text matches to help you decide what to read first — they do not determine eligibility, study quality or risk of bias. Edit the signals in the table below.

Criterion Label	Keywords (comma-separated, case-insensitive)	Rule type

Automated Quality Checklists

— run automatically on any paper with full text available

Once a paper's full text is retrieved (via PubMed Central or a free PDF), the app will automatically run quality checklists against the text. Each checklist item is keyword-scanned and pre-filled as a starting point — you can review and adjust each item manually inside the paper's expanded view. Checklists are matched to study type by default but can be changed per paper. The appropriate risk of bias tool is automatically selected and assessed alongside the matched quality checklist (e.g. RoB 2 for RCTs, ROBINS-I V2 for non-randomised intervention studies, Newcastle-Ottawa for cohort/case-control studies).

✓ CONSORT 2025 — RCTs ✓ PRISMA 2020 — Systematic Reviews & Meta-Analyses ✓ STROBE — Observational Studies ✓ CARE 2013 — Case Reports ✓ AGREE II — Clinical Guidelines ⚖ RoB 2 — RCT bias ⚖ ROBINS-I V2 — Non-randomised bias ⚖ Newcastle-Ottawa — Cohort/Case-Control ⚖ Jadad — RCT quality score

📄 How to Cite Evidentum Pro

If you use Evidentum Pro in your research or systematic review, please cite it using one of the formats below.

McNally, A., & Duckworth, A. (). Evidentum Pro: AI-powered systematic literature review tool (Version 1.0) [Software]. Retrieved from https://ankianaesthesia.evidentum.app/

↑ Import records Bring in Embase, CENTRAL or other exports (RIS or CSV) — deduplicated against your results, screened alongside them. Source ↓ Choose RIS/CSV file(s)

🔬

No results yet.

Everything your searches find lands here, deduplicated across databases and ready to screen.

Search History & Alerts

🔔 Toggle the bell on any search to save it as an alert. Check for new results any time.

No searches yet.

Every search you run is recorded here, with one-click re-run and new-result alerts.

📋 Standalone PDF Scanner

Upload any research paper PDF directly and run a quality checklist against it — no PubMed search needed. The app reads the PDF locally in your browser (nothing is sent to any server). Useful for spot-checking a specific paper or assessing papers outside a formal search.

📄

Drop a PDF here, or click to browse

PDF is read locally — nothing is uploaded to any server

Additional documents Appendices, supplementary data, statistical outputs

📋

Reporting Quality

How completely is the study reported?

▲

CONSORT 2025

30-item (42-row) checklist for transparent reporting of randomised trials.

RCT

PRISMA 2020

27-item checklist for systematic reviews and meta-analyses.

Systematic Review / Meta-Analysis

STROBE

22-item checklist for cohort, case-control, and cross-sectional studies.

Observational

CARE 2013

13-item checklist for completeness in clinical case reporting.

Case Report

AGREE II

23-item instrument across six domains for clinical practice guideline quality.

Clinical Guideline

⚖

Risk of Bias

How likely is the study to be biased?

▲

RoB 2

Cochrane risk of bias tool (2019). Five domains covering randomisation, deviations, missing data, measurement, and reporting.

RCT

ROBINS-I V2

Risk Of Bias In Non-randomised Studies of Interventions (V2, 2025). Seven bias domains adapted from RoB 2 for non-randomised intervention studies.

Non-randomised (interventions)

Newcastle-Ottawa Scale

Star-based scoring across selection, comparability, and outcome domains for cohort and case-control studies.

Cohort / Case-Control

Jadad Scale

5-point scale assessing randomisation, blinding, and withdrawals/dropouts in clinical trials.

RCT

📁 My Projects

✎ Keyword screening criteria

Tune the keyword sieve, then Apply to re-score the current results. Must contain flags a paper when any keyword is found; Must NOT contain flags it when none are. These criteria are saved with a project when you add papers to it.

Criterion	Keywords (comma-separated)	Rule

🧪 Validation harness

Compare the AI's risk-of-bias judgements for this project against a reference standard (e.g. the published review's consensus judgements). Export the template, fill the judgement columns (low / some concerns / high — or the ROBINS-I scale), import it back, and read the agreement report. The comparison never alters the appraisals.

Tool: ↑ Import reference CSV