Sparse TF-IDF Outperforms Dense Methods in SNOMED CT Linking
September 17, 2026
In a masked-concept recommendation benchmark for clinical notes, sparse TF-IDF achieved the best performance with a Recall@10 of 33.43%. This outperformed dense latent semantic analysis and retrieval-augmented hybrid approaches on the MIMIC-IV-Note dataset.
HOW THIS AFFECTS YOU
●
researcherEvaluate if density in embeddings is counterproductive for rare clinical entity linking tasks.
●
healthStandardizing clinical language may rely more on efficient sparse methods than complex dense embeddings for rare concepts.