Literature Constellations

Blueprint

Every tool in the lab is documented the same way, from the operational problem to the MLR questions a reviewer will ask.

1. 1

   Operational problem

   Medical affairs teams are asked what the literature looks like around a disease, a mechanism or a competitor: which themes dominate, where the trials are and what is still missing. A PubMed search returns a ranked list of hundreds of titles, and turning it into a picture of the field takes days of reading and tagging.

   The result is landscape reviews that are slow to start and evidence gaps that are noticed late, often by someone else.

2. 2

   Hypothesis

   If a search can be turned into a map in seconds, with topics named, study designs visible and empty corners flagged, teams will see the shape of a field before they read it. They can then plan reviews, publications and evidence generation around real gaps instead of impressions.

   The same map, played year by year, shows how a field grew, which helps with scientific narrative and with teaching.

3. 3

   Solution architecture

   • Relay (pubmed.php, PHP 7.4+): validated parameters, one fixed upstream (NCBI E-utilities esearch and efetch), no visitor headers or logs, NCBI's tool name and contact e-mail, calls spaced site-wide to respect the rate limit.
   • Cache of public records for 24 hours, so a repeated search does not hit NCBI again.
   • Engine (engine.js, in a Web Worker): TF-IDF vectors of titles, abstracts, MeSH and keywords; cosine similarity; exact t-SNE streamed to the screen; k-means++ constellations; class-based TF-IDF names; rule-based study designs; a layout check against chance.
   • Interface: canvas star map with pan and zoom, constellation cards, evidence matrix with gap cells, paper list, year player, CSV and PNG export, shareable links.
   • Tests: the relay end to end against a mock E-utilities serving real PubMed XML; layout quality and topic coherence on real records.
   • Installable PWA; a demo of real 2021 COVID-19 records works offline.
4. 4

   HCP and patient impact

   HCPs: better-targeted publications, reviews and medical education, because gaps in the evidence are found earlier.

   Patients: indirect. Evidence generation can be steered toward the questions that are still open.

   No patient-level data is processed at any point.

5. 5

   Regulatory and MLR considerations

   • Public bibliographic data only; no personal or HCP data. Searches are not stored with anything that identifies who made them.
   • For orientation, landscape analysis and teaching. It is not a systematic search and does not replace a documented search strategy.
   • No generative AI: the layout, names and designs are computed with classical, documented methods and can be traced to the records.
   • Data source attribution: PubMed, U.S. National Library of Medicine, through NCBI E-utilities. Literature Constellations is independent and not endorsed by NLM or NCBI.