Related Experiment Video
Updated: May 2, 2026
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
1.3K
Reinforcement learning improves LLM accuracy and reasoning in disease classification from radiology reports
Yishu Wei1,2, Yi Lin1, Adam Flanders3
1Department of Population Health Sciences, Weill Cornell Medicine, New York, NY, USA.
NPJ Digital Medicine
|April 30, 2026
Summary
This study introduces a two-stage method for accurate disease classification from radiology reports. The approach enhances classification accuracy and reasoning recall without sacrificing performance.
Area of Science:
- Medical Informatics
- Artificial Intelligence in Healthcare
- Natural Language Processing
Background:
- Accurate disease classification from radiology reports is crucial for clinical applications.
- Supervised fine-tuning (SFT) of large language models (LLMs) improves accuracy but can impair reasoning capabilities.
Purpose of the Study:
- To develop and evaluate a novel two-stage approach for disease classification in radiology reports.
- To enhance both classification accuracy and reasoning performance of LLMs.
Main Methods:
- A two-stage method involving supervised fine-tuning (SFT) on disease labels.
- Subsequent refinement using Group Relative Policy Optimization (GRPO) to optimize accuracy and format without direct reasoning supervision.
Main Results:
- SFT alone outperformed baseline methods in disease classification.
- GRPO further improved classification accuracy and significantly enhanced reasoning recall and comprehensiveness across three radiologist-annotated datasets.
Conclusions:
- The proposed two-stage approach effectively improves disease classification from radiology reports.
- This method balances accuracy optimization with enhanced reasoning capabilities, offering a promising advancement for clinical NLP applications.
