Reinforcement Learning Improves LLM Accuracy and Reasoning in Disease Classification From Radiology Reports

Authors

• Yishu Wei
• Yi Lin
• Adam Flanders, Thomas Jefferson UniversityFollow
• George Shih
• Yifan Peng

Document Type

Article

Publication Date

4-30-2026

Comments

This article is the author’s final published version in npj Digital Medicine, Volume 9, Issue 1, 2026, Article number 515.

The published version is available at https://doi.org/10.1038/s41746-026-02685-4. Copyright © The Author(s) 2026.

Abstract

Accurate disease classification from radiology reports is essential for many applications. While supervised fine-tuning (SFT) of lightweight LLMs improves accuracy, it can degrade reasoning. We propose a two-stage approach: SFT on disease labels followed by Group Relative Policy Optimization (GRPO) to refine predictions by optimizing accuracy and format without reasoning supervision. Across three radiologist-annotated datasets, SFT outperformed baselines and GRPO further improved classification and enhanced reasoning recall and comprehensiveness.

Recommended Citation

Wei, Yishu; Lin, Yi; Flanders, Adam; Shih, George; and Peng, Yifan, "Reinforcement Learning Improves LLM Accuracy and Reasoning in Disease Classification From Radiology Reports" (2026). Department of Radiology Faculty Papers. Paper 197.
https://jdc.jefferson.edu/radiologyfp/197

PubMed ID

42062541

Language

English