PPubMed5 Mar 2025Cited 1×

Biomedical Text Normalization through Generative Modeling

Jacob S. Berkowitz, Apoorva Srinivasan, Jose Miguel Acitores Cortina, Yasaman Fatapour, Nicholas P Tatonetti

Abstract

1.## INTRODUCTION The global transition to digital clinical records has led to a drastic increase in electronically stored unstructured data, making up around 80%(1) of the information within healthcare. This data has the potential to improve patient care, deepen our understanding of diseases, and support research. Yet the different ways medical terms are used in these records make it difficult to extract meaningful insights.(2) For instance, clinical notes often use ambiguous shorthand, inconsistent structure, and domain-specific vocabularies.(3) Text normalization is the mapping of natural l

A figure from Biomedical Text Normalization through Generative Modeling — fig. from the paper

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

Retrieval‑augmented generation (RAGnorm) outperforms other LLM‑based and baseline methods for biomedical text normalization, achieving the lowest mean shortest‑path errors and an 88% micro‑F1 on TAC‑2017.

What’s new

i.Developed four LLM‑based normalization strategies (Zero‑Shot Recall, Prompt Recall, Semantic Search, RAGnorm) and a TF‑IDF baseline.
ii.Created three SNOMED‑CT synonym datasets (oncology‑specific, random sample, high‑frequency) using GPT‑4 generated synonyms.
iii.Demonstrated that RAGnorm achieves the lowest mean shortest‑path lengths (0.21, 0.58, 0.90) across datasets and the highest micro‑F1 (88.01) on the TAC‑2017 task.
iv.Analyzed computational efficiency, showing RAGnorm reduces token usage compared to full prompting approaches.

How they did it

The authors implemented four LLM‑based normalization pipelines—Zero‑Shot Recall, Prompt Recall, Semantic Search, and Retrieval‑Augmented Generation (RAGnorm)—and a TF‑IDF string‑matching baseline. They generated synonym datasets for SNOMED‑CT terms using deterministic GPT‑4 outputs and evaluated performance with mean shortest‑path length and, for an external benchmark, micro‑F1 on the TAC‑2017 ADR dataset. Retrieval in RAGnorm used the GIST‑large‑Embedding‑v0 model to fetch relevant entries before GPT‑4 generated the final normalized term.

What they found

RAGnorm mean shortest‑path length: 0.21 (oncology), 0.58 (random sample), 0.90 (high‑frequency) – lowest among all methods.
RAGnorm micro‑F1 score of 88.01 on TAC‑2017 Task 4, surpassing other competitors.
TF‑IDF baseline mean shortest‑path lengths were 1.02, 1.98, and 2.37 for the three datasets, markedly higher than RAGnorm.

Caveats

✕RAGnorm requires two‑step retrieval and generation, incurring higher computational cost and token usage than simpler methods.
✕Dependence on proprietary LLM (GPT‑4) and external retrieval infrastructure may limit accessibility and real‑time clinical deployment.
✕Evaluation was limited to SNOMED‑CT and MedDRA datasets; broader clinical vocabularies were not tested.
✕Prompt Recall was unsustainable for large vocabularies, indicating scalability challenges for some approaches.

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

Similar papers in all four archives

The same idea searched in arXiv, PubMed, bioRxiv and chemRxiv at once, so a method can lead you to where it is being applied.

AarXiv4 found

PPubMed3 found

BbioRxiv4 found

CchemRxiv2 found

Searching the other archives…