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

2026-04-21

AI Summary

This work addresses the trade-off between classification accuracy and reasoning capability in radiology report disease classification, where supervised fine-tuning often improves accuracy at the expense of model interpretability. To reconcile this, the authors propose a two-stage approach: first, a lightweight large language model is fine-tuned with disease labels under supervision; subsequently, Group Relative Policy Optimization (GRPO) is applied to refine model outputs without requiring explicit reasoning annotations. This study presents the first application of GRPO to radiology text classification. Evaluated on three datasets annotated by radiologists, the method not only significantly outperforms baseline models in classification performance but also concurrently enhances reasoning recall and content comprehensiveness, achieving a synergistic improvement in both accuracy and reasoning quality.

Technology Category

• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Natural Language Processing: Safety and Robustness
• Machine Learning: Calibration & Uncertainty Quantification

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• disease classification
• radiology reports
• large language models
• reasoning
• accuracy

Innovation

• Reinforcement Learning
• Group Relative Policy Optimization
• Disease Classification
• Radiology Reports
• Reasoning Enhancement