G-CARL: Grounded Checklist-Aligned Reward Learning for Patient-Oriented Medical Report Interpretation

Analysis

Why This Paper Matters

Patient-oriented medical report interpretation is a critical yet underexplored area. Patients often receive complex medical reports without adequate explanation, leading to anxiety and misunderstanding. Existing medical vision-language tasks focus on generating factual reports for clinicians, but they do not address the need for accessible, patient-specific explanations. This paper introduces PMRI, a task that requires models to generate accurate and understandable interpretations based on a user's query and dialogue history. This dual requirement—factuality and patient communication—is challenging because these objectives are fundamentally different in verifiability and tightly coupled.

G-CARL addresses this challenge by providing structured supervision that balances factuality and patient alignment. The framework uses multi-source retrieval to verify atomic claims and instance-specific weighted checklists to ensure coverage of key points. This approach is significant because it moves beyond conventional supervised fine-tuning and holistic reinforcement learning, which often struggle to optimize these coupled objectives. The introduction of MMedReport, a real-world benchmark with clinician-designed evaluation, provides a standardized way to measure progress in this new task.

Technical Contributions

• PMRI Task Definition: Formalizes a new open-ended multimodal generation task that requires models to explain medical reports in accurate and accessible language based on user queries and dialogue history.
• G-CARL Framework: Combines multi-source retrieval for atomic claim verification with context-aware, instance-specific weighted checklists for response coverage. This provides structured supervision for factuality, user-demand satisfaction, and expression quality without constraining response diversity.
• MMedReport Benchmark: Constructs a real-world PMRI benchmark with a clinician-designed three-dimensional evaluation protocol, enabling comprehensive assessment of factuality, patient alignment, and expression quality.
• Reinforcement Learning Approach: Uses grounded, checklist-aligned rewards to optimize the model, addressing the challenge of jointly optimizing verifiable factuality and subjective patient communication.

Results

The paper reports that G-CARL consistently outperforms existing post-training baselines in overall quality, claim-level precision, and checklist recall. Specifically, the framework improves claim-level precision, indicating more accurate medical statements, and checklist recall, suggesting better coverage of important points. Clinician pairwise preference evaluations further confirm that G-CARL produces interpretations that are more accurate and better aligned with patient needs. These results demonstrate the effectiveness of the grounded, checklist-aligned reward learning approach in balancing factuality and patient communication.

Significance

This work has significant implications for medical AI and human-computer interaction. By introducing a task that prioritizes patient understanding, it encourages the development of models that are not only clinically accurate but also empathetic and accessible. The G-CARL framework provides a generalizable approach for optimizing multiple coupled objectives in multimodal generation, which could be applied to other domains where factuality and user-centric communication are both critical. The MMedReport benchmark offers a valuable resource for future research, potentially driving progress in personalized medical communication and patient education.