Reinforcement learning improves LLM accuracy and reasoning in disease classification from radiology reports.

Key Points

• This research aims to improve the accuracy and reasoning of large language models (LLMs) in disease classification from radiology reports.
• Implemented a two-stage approach combining supervised fine-tuning on disease labels and Group Relative Policy Optimization.
• Evaluated across three radiologist-annotated datasets to assess performance against baseline models.
• Analyzed both accuracy and reasoning enhancements without direct reasoning supervision.
• Supervised fine-tuning led to significant performance improvement over baseline models.
• Group Relative Policy Optimization further enhanced classification accuracy and reasoning recall.
• Overall, the combination yielded comprehensiveness in disease classification from radiology reports.