Project 9: Health-literacy gap as embedding distance¶
Mentors: To be announced
Problem: "This topic is hard to explain to patients" is a judgment call today, with no computable score behind it.
Context: Built on MedlinePlus consumer health articles paired with matched PubMed abstracts pulled via the NCBI E-utilities API (esearch/efetch, free, no key required for light use) on the same condition, anchored to ML4LLM Ch.3 · proj13: Word similarity via distance and cosine (helper).
Goals: Can the plain-language vs. clinical-language gap for the same disease be measured as embedding distance, giving a computable health-literacy score?
Deliverables: A notebook that pairs MedlinePlus consumer articles with a PubMed abstract on the same condition, embeds both, computes distance/cosine similarity, and ranks condition pairs by embedding distance to identify which topics have the largest plain-language-to-clinical-language gap.
Showcase: TBD
References:
- Language Modeling by Clustering with Word Embeddings for Text Readability Assessment: uses embedding-space distance as a readability signal, the same core idea this project applies to the plain-language-vs-clinical-language gap.