Four million anonymized Italian emergency-room records are now freely available for medical AI research.
A team of researchers has published eCream-MedCorpus, described as the largest freely available dataset of Italian clinical notes in existence. The corpus draws from Emergency Department records across Italian hospitals, covering multiple phases of a patient's stay. A smaller annotated subset — roughly 6,000 notes — was hand-labeled by clinical experts using a 132-item structured form covering two specific conditions: dyspnea and loss of consciousness. Labels span numerical, categorical, binary, and mixed value types. Anonymization happened on-site before any data left the hospitals.
Most publicly available medical language datasets skew heavily toward English, which makes it harder to build or evaluate clinical AI tools for systems operating in other languages. A dataset this size, built from real emergency-room records and annotated by actual clinicians, gives Italian researchers something to train against rather than adapt around. The paper also proposes a new benchmark task — CRF-filling, or automatically completing structured case report forms from raw notes — and runs zero-shot baselines using Gemma-27B and MedGemma-27B.
The annotated subset is notably imbalanced, which the researchers acknowledge — a realistic reflection of how rare certain clinical presentations are, but also a known headache for model training. Whether the dataset accelerates genuinely useful clinical tools or mostly benchmarks that look good on paper will depend on what the research community does with it next.