Generation of Bilingual Synthetic Clinical Notes for Realistic Data Augmentation

David Seung U Lee, Seeun Park, Seoyoon Jang, Sunyoung Lee, Chaeyoung Chang, Sungwook Choi, Howard Lee
Proceedings of the 7th Conference on Health, Inference, and Learning, PMLR 333:834-879, 2026.

Abstract

Synthetic clinical notes offer a promising solution to data scarcity and privacy constraints in clinical natural language processing. However, existing generation approaches often prioritize semantic accuracy while not adequately reproducing the linguistic and structural (i.e., surface) characteristics of real-world clinical documentation, limiting their utility for downstream clinical tasks. In this study, we propose an expert-informed prompt with feedback-loop generation framework to improve the fidelity of synthetic clinical notes across both semantic and surface-level dimensions. Using individual case safety reports from FAERS, we formulated synthetic note generation as a controlled text generation task conditioned on adverse drug reaction descriptions and clinical narratives. We evaluated the performance of the proposed approach by comparing it with other generation strategies (in-context learning and multi-agent generation) and prompting methods (base and expert-informed) under a unified experimental condition. Generation quality was assessed using embedding-based semantic similarity, surface-level statistical and distributional metrics, and blinded human evaluation. The feedback-loop generation framework achieved superior performance across semantic (mean clinical BERTScore = 0.885) and surface-level distributional metrics (token-level Jensen-Shannon divergence = 0.344), producing synthetic clinical notes that more closely resembled real-world clinical notes than other approaches. Expert-informed prompting further improved semantic fidelity and lexical diversity.

Cite this Paper

BibTeX
@InProceedings{pmlr-v333-lee26b, title = {Generation of Bilingual Synthetic Clinical Notes for Realistic Data Augmentation}, author = {Lee, David Seung U and Park, Seeun and Jang, Seoyoon and Lee, Sunyoung and Chang, Chaeyoung and Choi, Sungwook and Lee, Howard}, booktitle = {Proceedings of the 7th Conference on Health, Inference, and Learning}, pages = {834--879}, year = {2026}, editor = {Healey, Elizabeth and Fries, Jason and Pollard, Tom and Tang, Shengpu and Zink, Anna and Hartvigsen, Tom and Agrawal, Monica and Finlayson, Sam and Glicksberg, Benjamin and Beaulieu-Jones, Brett and Wang, Kai and Fontalvo, Daseyra and Sarker, Tasmie and Chen, Irene and Alsentzer, Emily}, volume = {333}, series = {Proceedings of Machine Learning Research}, month = {29--30 Jun}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v333/main/assets/lee26b/lee26b.pdf}, url = {https://proceedings.mlr.press/v333/lee26b.html}, abstract = {Synthetic clinical notes offer a promising solution to data scarcity and privacy constraints in clinical natural language processing. However, existing generation approaches often prioritize semantic accuracy while not adequately reproducing the linguistic and structural (i.e., surface) characteristics of real-world clinical documentation, limiting their utility for downstream clinical tasks. In this study, we propose an expert-informed prompt with feedback-loop generation framework to improve the fidelity of synthetic clinical notes across both semantic and surface-level dimensions. Using individual case safety reports from FAERS, we formulated synthetic note generation as a controlled text generation task conditioned on adverse drug reaction descriptions and clinical narratives. We evaluated the performance of the proposed approach by comparing it with other generation strategies (in-context learning and multi-agent generation) and prompting methods (base and expert-informed) under a unified experimental condition. Generation quality was assessed using embedding-based semantic similarity, surface-level statistical and distributional metrics, and blinded human evaluation. The feedback-loop generation framework achieved superior performance across semantic (mean clinical BERTScore = 0.885) and surface-level distributional metrics (token-level Jensen-Shannon divergence = 0.344), producing synthetic clinical notes that more closely resembled real-world clinical notes than other approaches. Expert-informed prompting further improved semantic fidelity and lexical diversity.} }
Endnote
%0 Conference Paper %T Generation of Bilingual Synthetic Clinical Notes for Realistic Data Augmentation %A David Seung U Lee %A Seeun Park %A Seoyoon Jang %A Sunyoung Lee %A Chaeyoung Chang %A Sungwook Choi %A Howard Lee %B Proceedings of the 7th Conference on Health, Inference, and Learning %C Proceedings of Machine Learning Research %D 2026 %E Elizabeth Healey %E Jason Fries %E Tom Pollard %E Shengpu Tang %E Anna Zink %E Tom Hartvigsen %E Monica Agrawal %E Sam Finlayson %E Benjamin Glicksberg %E Brett Beaulieu-Jones %E Kai Wang %E Daseyra Fontalvo %E Tasmie Sarker %E Irene Chen %E Emily Alsentzer %F pmlr-v333-lee26b %I PMLR %P 834--879 %U https://proceedings.mlr.press/v333/lee26b.html %V 333 %X Synthetic clinical notes offer a promising solution to data scarcity and privacy constraints in clinical natural language processing. However, existing generation approaches often prioritize semantic accuracy while not adequately reproducing the linguistic and structural (i.e., surface) characteristics of real-world clinical documentation, limiting their utility for downstream clinical tasks. In this study, we propose an expert-informed prompt with feedback-loop generation framework to improve the fidelity of synthetic clinical notes across both semantic and surface-level dimensions. Using individual case safety reports from FAERS, we formulated synthetic note generation as a controlled text generation task conditioned on adverse drug reaction descriptions and clinical narratives. We evaluated the performance of the proposed approach by comparing it with other generation strategies (in-context learning and multi-agent generation) and prompting methods (base and expert-informed) under a unified experimental condition. Generation quality was assessed using embedding-based semantic similarity, surface-level statistical and distributional metrics, and blinded human evaluation. The feedback-loop generation framework achieved superior performance across semantic (mean clinical BERTScore = 0.885) and surface-level distributional metrics (token-level Jensen-Shannon divergence = 0.344), producing synthetic clinical notes that more closely resembled real-world clinical notes than other approaches. Expert-informed prompting further improved semantic fidelity and lexical diversity.
APA
Lee, D.S.U., Park, S., Jang, S., Lee, S., Chang, C., Choi, S. & Lee, H.. (2026). Generation of Bilingual Synthetic Clinical Notes for Realistic Data Augmentation. Proceedings of the 7th Conference on Health, Inference, and Learning, in Proceedings of Machine Learning Research 333:834-879 Available from https://proceedings.mlr.press/v333/lee26b.html.

Related Material