MedCaseAgent is an end‑to‑end agentic reasoning paradigm that converts unstructured clinical notes into publication‑quality reports for rare diseases. The system follows a hierarchical design: it first mines patient‑level data, then performs iterative, tool‑integrated reasoning to synthesize evidence. By dynamically querying medical literature and knowledge bases, the agent preserves clinical fidelity and gauges case novelty. The framework is modular, allowing plug‑and‑play support for various backbone large language models. To accelerate research, we release the MedCase‑150K multimodal case‑report corpus and the temporally held‑out MedCase‑Bench evaluation suite. An open‑source model fine‑tuned from Qwen3.5‑9B is also provided, optimized specifically for this task. Across both LLM‑judged and rule‑based metrics, MedCaseAgent outperforms existing general‑purpose and specialized medical LLMs and agents. Blind expert reviews and ablation studies further confirm its clinical relevance and evidence grounding.
Review: The system demonstrates feasibility and advantage in generating rare‑disease case reports.