EDEN: 4M Italian clinical notes dataset to boost medical LLMs

Researchers from Italian institutions have unveiled EDEN, a massive corpus of ~4 million clinical notes sourced from emergency departments across Italian hospitals. The dataset is fully anonymized and covers diverse phases of patient care. A hand-annotated subset of about 6,000 notes focuses on two conditions—dyspnea and loss of consciousness—using a structured Case Report Form with 132 items (numerical, categorical, binary, and mixed). Multiple clinicians iteratively revised annotations to resolve ambiguities, producing a rich but imbalanced resource. The team also proposes CRF-filling as a structured information extraction benchmark, providing zero-shot baselines from Google's Gemma-27B and MedGemma-27B models.

EDEN is the largest freely available corpus of clinical notes for Italian, filling a critical gap in non-English medical NLP. By offering both raw notes and structured annotations, it enables development of LLMs for tasks like triage support, discharge summarization, and clinical decision-making. The dataset's release is expected to accelerate research into Italian-language medical AI, with potential applications across European healthcare systems. Researchers describe their data collection protocol, on-site anonymization pipeline, corpus statistics, and annotation scheme in the accompanying paper.