An open prototype for cited, evidence-graded answers to nursing questions.

Ask a clinical question in plain language. EvidenceRN searches PubMed, screens out retracted records, labels each source by evidence type (Melnyk & Fineout-Overholt hierarchy), and writes a short synthesized answer where every claim links to a numbered source. Follow-up questions reuse the same sources. A "deep read" pulls open-access full text from PubMed Central to refine the answer.

Status: prototype (v0.5.3). EvidenceRN is a project of AI + Nurse Academy, built by a nurse practitioner to show what a transparent, open evidence assistant for nurses looks like. It is not a clinical decision support product. See LIMITATIONS.md before using it for anything beyond exploration and education.

Nurses need fast, trustworthy answers in every setting where they practice, from hospitals and clinics to schools, homes, and communities. Commercial products are arriving in this space with proprietary evidence bases and institutional contracts. EvidenceRN takes a different path on purpose: public literature, visible search strategy, labeled evidence levels, retraction screening, and a citation check on every answer, all in a single readable file anyone can inspect, run, and improve.

The code is open under Apache 2.0. The name EvidenceRN belongs to the project founder (see NOTICE).

• Converts a conversational question into a PubMed Boolean query, with the exact query shown to the user
• Filters by evidence type (best available, systematic reviews, guidelines, RCTs, or any), recency, and nursing focus, and relaxes filters one axis at a time when strict searches return nothing
• Retrieves a wider candidate pool than requested, ranks by evidence level, and keeps the strongest set
• Excludes retracted publications and retraction notices using PubMed publication types and CommentsCorrections links; flags expressions of concern, errata, and linked updates on source cards
• Generates a three-section answer (bottom line, what the evidence shows, practice implications) with bracketed citations, framed for a generalist RN scope that is stated under every answer, and required to separate "the sources did not study this" from "the sources found no difference"
• Runs a heuristic citation check on every answer and follow-up: citation numbers outside the source list and substantive sentences without a citation get flagged
• Deep read: retrieves PubMed Central full text for top Level I sources, reads sections in ranked order (results, then discussion and conclusions, then methods, then introduction) within a 12,000 character budget, and revises the answer, with a "what the full text added" section and a comparison to the abstract-only answer
• Surfaces a hand-curated index of gray literature (CDC, AHRQ, NICE, ISMP, and others) by topic, clearly marked as not read by the synthesis engine
• Shows an urgent-situation banner when a question describes a potentially acute event
• Discloses that no institutional policy is connected

Two files do the work.

public/index.html is the whole user interface and client logic, dependency-free. It calls NLM E-utilities directly from the browser for search, abstracts, and PMC full text.

server.js is a zero-dependency Node server. It serves the UI and exposes one endpoint, POST /api/ask, which forwards a synthesis prompt to the Claude API using the key in .env. The key never reaches the browser.

Swapping the language model means editing one function in server.js. Swapping or adding an evidence source means adding a function alongside pubmedSearch in index.html.

Requirements: Node.js 18 or newer, and an Anthropic API key from https://console.anthropic.com for the synthesis features. Search works without a key.

git clone https://github.com/EvidenceRN/evidencern.git
cd evidencern
cp .env.example .env     # paste your key after ANTHROPIC_API_KEY=
npm start                # or: node server.js

Open http://localhost:3000.

The project deploys as-is to Render, Fly.io, Railway, or any host running Node: start command node server.js, environment variable ANTHROPIC_API_KEY. Before exposing it publicly, add authentication or invite codes and per-user rate limits so strangers cannot spend your API budget. Neither exists yet; see the roadmap.

Each search makes one or two model calls (query interpretation on longer questions, plus the synthesis). Follow-ups and deep reads add one call each. Ten abstracts per synthesis costs a few cents per question on a Sonnet-class model and less on Haiku. Set ANTHROPIC_MODEL in .env to change models.

• No patient data is stored. Users should keep patient identifiers out of the search box; the UI does not yet enforce this.
• Evidence labels reflect study design inferred from PubMed metadata. No risk-of-bias assessment or GRADE rating is performed.
• The synthesis reads abstracts only, unless deep read succeeds. Abstracts omit methods detail, harms, and limitations.
• The footer disclaimer in the UI states these limits to every user. Keep it in any derivative.

Contributions are welcome, and the project is actively looking for two or three maintainers with engineering experience. Read CONTRIBUTING.md for how the project works, what help is most needed, and how maintainership is decided.

Apache License 2.0. See LICENSE and NOTICE.

EvidenceRN is a project of AI + Nurse Academy, which offers nurse-led AI education for nurses and other clinicians. Project site: https://evidencern.com

Sources come from PubMed and PubMed Central, services of the U.S. National Library of Medicine. NLM does not endorse this software. Evidence-level labels follow Melnyk, B. M., & Fineout-Overholt, E., Evidence-Based Practice in Nursing & Healthcare.

The v0.5.3 changes (section-aware deep read, longer abstract windows, a separate label for trials whose PubMed metadata states no control group, the rule separating no evidence found from evidence of no effect, and the stated scope assumption) respond to a clinician code review by Sarah Curtright, DNP, MBA, in October 2026.