Open-Ended Clinical Text Generation for Acute Care: Applying Reinforcement Learning with Clinically Grounded Rewards

Minjia Wang, Luyang Luo, Sung Eun Kim, Fang Cao, David A Kim, Pranav Rajpurkar
Proceedings of the 7th Conference on Health, Inference, and Learning, PMLR 333:966-984, 2026.

Abstract

Acute care clinicians generate critical clinical text—diagnoses, treatment plans, discharge instructions—under time pressure where errors can be life-threatening. Large proprietary AI models raise privacy concerns, while smaller models lack clinical quality. We extend reinforcement learning with verifiable rewards (RLVR) to open-ended clinical text generation using two generalizable reward patterns: equivalence-based rewards for medical synonymy and diagnosis matching, as well as rubric-based rewards for multi-dimensional quality assessment. Using group relative policy optimization, we trained compact 7–8 billion parameter models on diagnosis generation (MIMIC-III), discharge instructions (DischargeMe), and treatment planning (MTSamples). Trained models achieve clinical quality across tasks (best results: F1 0.48, 4.28/5.0, 4.47/5.0 respectively), matching or surpassing the performance of large proprietary GPT-based models, while enabling on-premise deployment, sub-second inference, and full privacy. Physician review confirmed superior content comprehensiveness and fewer dangerous errors versus base models. This demonstrates a practical pathway for deploying clinical text generation in acute care with generalizable reward design patterns.

Cite this Paper

BibTeX
@InProceedings{pmlr-v333-wang26d, title = {Open-Ended Clinical Text Generation for Acute Care: Applying Reinforcement Learning with Clinically Grounded Rewards}, author = {Wang, Minjia and Luo, Luyang and Kim, Sung Eun and Cao, Fang and Kim, David A and Rajpurkar, Pranav}, booktitle = {Proceedings of the 7th Conference on Health, Inference, and Learning}, pages = {966--984}, 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/wang26d/wang26d.pdf}, url = {https://proceedings.mlr.press/v333/wang26d.html}, abstract = {Acute care clinicians generate critical clinical text—diagnoses, treatment plans, discharge instructions—under time pressure where errors can be life-threatening. Large proprietary AI models raise privacy concerns, while smaller models lack clinical quality. We extend reinforcement learning with verifiable rewards (RLVR) to open-ended clinical text generation using two generalizable reward patterns: equivalence-based rewards for medical synonymy and diagnosis matching, as well as rubric-based rewards for multi-dimensional quality assessment. Using group relative policy optimization, we trained compact 7–8 billion parameter models on diagnosis generation (MIMIC-III), discharge instructions (DischargeMe), and treatment planning (MTSamples). Trained models achieve clinical quality across tasks (best results: F1 0.48, 4.28/5.0, 4.47/5.0 respectively), matching or surpassing the performance of large proprietary GPT-based models, while enabling on-premise deployment, sub-second inference, and full privacy. Physician review confirmed superior content comprehensiveness and fewer dangerous errors versus base models. This demonstrates a practical pathway for deploying clinical text generation in acute care with generalizable reward design patterns.} }
Endnote
%0 Conference Paper %T Open-Ended Clinical Text Generation for Acute Care: Applying Reinforcement Learning with Clinically Grounded Rewards %A Minjia Wang %A Luyang Luo %A Sung Eun Kim %A Fang Cao %A David A Kim %A Pranav Rajpurkar %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-wang26d %I PMLR %P 966--984 %U https://proceedings.mlr.press/v333/wang26d.html %V 333 %X Acute care clinicians generate critical clinical text—diagnoses, treatment plans, discharge instructions—under time pressure where errors can be life-threatening. Large proprietary AI models raise privacy concerns, while smaller models lack clinical quality. We extend reinforcement learning with verifiable rewards (RLVR) to open-ended clinical text generation using two generalizable reward patterns: equivalence-based rewards for medical synonymy and diagnosis matching, as well as rubric-based rewards for multi-dimensional quality assessment. Using group relative policy optimization, we trained compact 7–8 billion parameter models on diagnosis generation (MIMIC-III), discharge instructions (DischargeMe), and treatment planning (MTSamples). Trained models achieve clinical quality across tasks (best results: F1 0.48, 4.28/5.0, 4.47/5.0 respectively), matching or surpassing the performance of large proprietary GPT-based models, while enabling on-premise deployment, sub-second inference, and full privacy. Physician review confirmed superior content comprehensiveness and fewer dangerous errors versus base models. This demonstrates a practical pathway for deploying clinical text generation in acute care with generalizable reward design patterns.
APA
Wang, M., Luo, L., Kim, S.E., Cao, F., Kim, D.A. & Rajpurkar, P.. (2026). Open-Ended Clinical Text Generation for Acute Care: Applying Reinforcement Learning with Clinically Grounded Rewards. Proceedings of the 7th Conference on Health, Inference, and Learning, in Proceedings of Machine Learning Research 333:966-984 Available from https://proceedings.mlr.press/v333/wang26d.html.

Related Material