Reproduce “MedSyn: LLM-based Synthetic Medical Text Generation Framework” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gleb Kumichev, Pavel Blinov, Yulia Kuzkina, Vasily Goncharov, Galina Zubkova, Nikolai Zenovkin, Aleksei Goncharov, Andrey Savchenko
MedSyn, a framework developed by researchers from MIL Team and Sber AI Lab, generates synthetic Russian medical texts by integrating large language models with a Medical Knowledge Graph. The approach improved ICD-code classification accuracy by up to 17.8% for specific vital codes and resulted in synthetic texts that human evaluators found challenging to distinguish from real notes.
Key insights
- Integrating structured medical knowledge from an MKG is effective in guiding LLMs to produce factually accurate and coherent synthetic medical texts, mitigating the issue of hallucinations.
- Synthetic data generation can successfully address data scarcity and imbalance in medical datasets, leading to improved performance of downstream NLP models, especially for underrepresented conditions.
- Fine-tuning general-purpose LLMs on instruction-following datasets derived from medical knowledge graphs and real-world clinical tasks enables robust medical text generation for low-resource languages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.