About

Clinicle launched in January 2023 and was rebuilt in 2026.

2023

A word2vec model from Gary Weissman's group at Penn LDI trained on open-access medical case reports, cleaned through ScispaCy's UMLS entity linker. React frontend, Python on Heroku, Firebase storage.

2026

Scoring runs on SapBERT — a PubMedBERT model contrastively fine-tuned on UMLS synonym pairs, so cosine similarity tracks clinical concept identity rather than general text overlap. 768-dimensional vectors, rescaled so a random guess lands near 0 and a true synonym near 100. Vocab is a curated union of ~1,000 puzzle diagnoses, ~265k MeSH concepts, a hand-picked list of common medical abbreviations (MI, LAD, ITP, COPD…) each embedded with its expansion, and common English for cold guesses. Static Next.js on Vercel; no backend, no login.

Source on GitHub.