G-CARL: Grounded Checklist-Aligned Reward Learning for Patient-Oriented Medical Report Interpretation
G-CARL employs retrieval-grounded claim verification and checklist-aligned reinforcement learning to enhance medical report interpretation accuracy and patient-centricity.
Key Findings
Methodology
G-CARL adopts a multi-objective reinforcement learning framework based on GRPO, integrating retrieval-based claim verification and clinician-designed weighted checklists. The system first generates candidate responses, then decomposes responses into atomic claims, which are verified against a large-scale medical datastore (including drug instructions, textbooks, and guidelines). A multimodal verifier assesses claim support and relevance, providing localized rewards. Simultaneously, a case-specific checklist is constructed via large multimodal language models refined by clinicians, with each item weighted by clinical importance. The overall reward combines factuality, demand satisfaction, and expression quality, guiding the policy model to optimize these aspects simultaneously.
Key Results
- On the MMedReport dataset, G-CARL outperforms supervised fine-tuning and single-reward baselines, with a 0.3 point increase in subjective score, 0.77% higher claim precision, and 6.71% higher checklist recall. It also achieves significant improvements in factual correctness and user satisfaction metrics.
- External CMB evaluation shows G-CARL improves medical QA accuracy by 0.63 points and attains a professionalism score of 3.61, demonstrating its robustness and generalization in real-world scenarios.
- Ablation studies reveal that retrieval-grounded claim verification and personalized checklists are critical for performance gains, while static or absent retrieval mechanisms lead to notable drops in accuracy and coverage.
Significance
This work addresses the critical challenge of balancing factual accuracy and personalized communication in medical report interpretation. By integrating structured retrieval and dynamic checklist supervision within a reinforcement learning paradigm, G-CARL advances the reliability and patient-centricity of AI-generated medical explanations. Its capacity to produce trustworthy, comprehensive, and understandable reports has profound implications for clinical practice, patient education, and telemedicine, potentially transforming how medical information is conveyed and understood. The structured reward design offers a scalable solution to improve AI safety and interpretability in healthcare.
Technical Contribution
The paper introduces a novel multi-source retrieval grounded claim verification mechanism, enabling localized factual supervision. It further proposes a dynamic, clinician-refined weighted checklist system that explicitly supervises demand satisfaction and expression quality without constraining diversity. The integration of these components into a multi-objective reinforcement learning framework, supported by a formalized reward function, sets a new standard for trustworthy medical language models. The approach effectively mitigates hallucination issues and enhances model adaptability to case-specific needs, offering a theoretically sound and practically effective solution for medical AI interpretability.
Novelty
This is the first work to combine retrieval-based claim verification with case-specific, weighted checklists in a reinforcement learning setting for medical report explanation. Unlike prior approaches that rely on static rubrics or holistic scoring, G-CARL dynamically constructs and evaluates personalized checklists, enabling fine-grained supervision aligned with clinical priorities. Its claim-level factual verification mechanism, grounded in multi-source evidence retrieval, significantly improves factual accuracy and interpretability, marking a substantial step forward in trustworthy medical AI.
Limitations
- 依赖于高质量、全面的医学知识库，知识库的维护和更新成本较高，且在罕见或新兴病例中验证效果有限。
- 检查表设计虽由临床专家指导，但仍存在主观偏差，自动化和客观化程度不足，未来需引入更智能的自动化评估机制。
- 训练过程计算成本较大，硬件需求高，实际部署时可能面临性能瓶颈，需优化模型效率和推理速度。
Future Work
未来将探索自动化更新医学知识库的方法，结合实时临床数据增强模型的泛化能力。同时，计划引入患者反馈机制，持续优化检查表设计，实现个性化、动态化的解释生成。还将进行临床试验验证模型在真实医疗场景中的效果，推动模型的临床应用落地，提升医疗AI的可信度和实用性。
AI Executive Summary
In recent years, the demand for personalized medical report explanations has surged, driven by the need for patients to better understand complex diagnostic information. Existing models often struggle to balance factual accuracy with user-specific communication, leading to either unreliable outputs or overly generic responses. Traditional supervised fine-tuning tends to overfit to reference answers, limiting response diversity and adaptability. To address these issues, this study introduces G-CARL, a novel reinforcement learning framework that integrates retrieval-grounded claim verification with clinician-designed, weighted checklists. G-CARL decomposes responses into atomic claims, verifies each claim against a comprehensive medical datastore, and constructs personalized checklists to guide response quality. This multi-objective approach ensures that generated explanations are both factually accurate and aligned with individual patient concerns. Extensive experiments on the MMedReport dataset demonstrate that G-CARL significantly outperforms baseline methods, achieving higher claim-level precision and checklist recall, and producing more trustworthy, patient-oriented explanations. External evaluations on the CMB benchmark further confirm its robustness and generalization capabilities. The framework’s ability to incorporate structured, case-specific supervision marks a substantial advance in medical AI interpretability, promising safer and more effective patient communication. Looking ahead, integrating automated knowledge base updates and real-time patient feedback could further enhance the system’s adaptability and clinical utility, paving the way for widespread deployment in healthcare settings.
Deep Dive
Abstract
Personalized interpretation of medical reports has emerged as an increasingly important need among patients. Addressing this need requires both evidence-grounded medical factuality and context-dependent patient communication, yet existing medical vision-language tasks do not adequately capture these dual requirements. To bridge this gap, we introduce Patient-oriented Medical Report Interpretation (PMRI), a novel open-ended multimodal generation task that requires models to explain medical reports in accurate and accessible language based on a user's query and dialogue history. These two objectives differ fundamentally in their verifiability, yet remain tightly coupled, making them difficult to optimize jointly under conventional supervised fine-tuning and holistic reinforcement learning paradigms. To address this challenge, we propose G-CARL, a grounded, checklist-aligned reinforcement learning framework that combines multi-source retrieval for atomic claim verification with context-aware, instance-specific weighted checklists for response coverage, providing structured supervision for factuality, user-demand satisfaction, and expression quality without constraining response diversity. We further construct MMedReport, a real-world PMRI benchmark, along with a clinician-designed three-dimensional evaluation protocol. Extensive experiments demonstrate that G-CARL consistently outperforms existing post-training baselines in overall quality, claim-level precision, and checklist recall. Pairwise preference evaluation by clinicians further confirms that G-CARL produces interpretations that are more accurate and better aligned with patient needs.