MedSyn: LLM-based synthetic medical text generation framework

P. 215–230.

Kumichev G., Blinov P., Kuzkina Y., Goncharov V., Zubkova G., Zenovkin N., Goncharov A., Savchenko A.

Generating synthetic text addresses the challenge of data availability in privacy-sensitive domains such as healthcare. This study explores the applicability of synthetic data in real-world medical settings. We introduce MedSyn, a novel medical text generation framework that integrates large language models with a Medical Knowledge Graph (MKG). We use MKG to sample prior medical information for the prompt and generate synthetic clinical notes with GPT-4 and fine-tuned LLaMA models. We assess the benefit of synthetic data through application in the ICD code prediction task. Our research indicates that synthetic data can increase the classification accuracy of vital and challenging
codes by up to 17.8% compared to settings without synthetic data. Furthermore, to provide new data for further research in the healthcare domain, we present the largest open-source synthetic dataset of clinical
notes for the Russian language, comprising over 41k samples covering 219 ICD-10 codes.

Keywords: Synthetic data, Clinical note generation, ICD code prediction

In book

Machine Learning and Knowledge Discovery in Databases. Applied Data Science Track. European Conference, ECML PKDD 2024, Vilnius, Lithuania, September 9–13, 2024, Proceedings, Part X. LNCS, volume 14950

Cham: Springer, 2024.