MedCaseAgent: From Multimodal Clinical Evidence to Rare Disease Case Reports

[arXiv] Hongkang Chu, Yifan Shen, Tianjiao Yu and others propose MedCaseAgent, an end-to-end agentic reasoning framework that can synthesize publication-quality case reports from unstructured clinical records. According to the paper, the framework adopts a hierarchical structure, combining patient-level data mining with iterative tool-integrated reasoning, and maintains clinical fidelity and assesses case novelty by dynamically querying medical literature and knowledge bases; the framework is modular and supports plug-and-play use of a variety of base large models. The team also releases MedCase-150K, a large-scale multimodal clinical case report corpus, and MedCase-Bench, a temporal holdout evaluation set, and open-sources a model post-trained on Qwen3.5-9B and optimized for this task. The study states that under large-model judging and rule-based evaluation, MedCaseAgent outperforms existing general-purpose and specialized medical large models and agents, and blinded expert review and ablation experiments also support its clinical relevance and evidence grounding.