EDEN: A Crucial Corpus for LLMs in the Italian Healthcare Sector

EDEN (Emergency Department Electronic Notes) represents a new and significant large-scale corpus of clinical notes, entirely generated within the emergency departments of Italian hospitals. This resource, in its current version, comprises approximately 4 million fully anonymized clinical notes, covering various phases of patient care during their stay in the emergency department. The availability of such a vast and specific dataset for the Italian context is crucial for the development of Large Language Models (LLMs) capable of operating effectively in the medical field, a sector where precision and contextual understanding are paramount.

The creation of high-quality linguistic resources for specific domains and non-English languages is a prerequisite for the widespread adoption of artificial intelligence. EDEN aims to fill a significant gap, providing a robust foundation for the research and application of LLMs in the Italian healthcare landscape, with direct implications for diagnostics, patient management, and operational efficiency.

Technical and Methodological Details of the EDEN Corpus

In addition to the extensive collection of anonymized notes, EDEN includes a subset of about six thousand notes that have been manually annotated by clinical experts. This annotation process was carried out using a structured Case Report Form (CRF), containing 132 items relevant to two common clinical situations in emergency departments: dyspnea and loss of consciousness. Items can assume numerical values (such as blood saturation), categorical values (such as level of consciousness), binary values (such as the presence of traumas), or mixed value types.

The involvement of multiple clinicians and iterative revision helped resolve ambiguities in item formulation, creating a richly structured resource, albeit with some inherent imbalance typical of real-world data. The dataset also describes the data collection protocol, the on-site anonymization pipeline, corpus statistics, and the annotation scheme. Furthermore, CRF-filling is proposed as a novel structured information extraction benchmark, with zero-shot baselines obtained from Gemma-27B and MedGemma-27B, providing a reference point for future developments.

Implications for On-Premise Deployments and Data Sovereignty

The aspect of on-site anonymization is particularly relevant for organizations prioritizing data sovereignty and on-premise deployments. Local management of the anonymization process ensures greater control over sensitive data, meeting stringent compliance requirements such as GDPR. For CTOs, DevOps leads, and infrastructure architects evaluating self-hosted alternatives to the cloud for AI/LLM workloads, a corpus like EDEN, freely available and with data already anonymized locally, significantly reduces the risks associated with transferring and managing health information in the public cloud.

This approach supports the creation of air-gapped or hybrid environments, where models can be trained or used for inference while keeping data within corporate or national boundaries. Such a strategy not only strengthens security and compliance but can also optimize the Total Cost of Ownership (TCO) in the long term, avoiding the recurring and unpredictable costs associated with processing large volumes of sensitive data on external cloud platforms.

Future Prospects and AI-RADAR Context

The EDEN dataset stands as the largest freely available corpus of clinical notes for the Italian language, filling a significant gap in the landscape of resources for LLM research and development. Its availability is an enabling factor for innovation in concrete medical applications, from assisted diagnostics to patient management, accelerating the adoption of AI solutions in the healthcare sector.

For those evaluating on-premise LLM deployments, the existence of a high-quality, locally controlled dataset like EDEN offers a strategic advantage, allowing for the development of AI solutions without compromising data privacy or sovereignty. AI-RADAR, in its commitment to analyzing the trade-offs between self-hosted and cloud solutions, emphasizes how resources like EDEN are fundamental for building robust and compliant local stacks, providing the foundation for responsible and controlled artificial intelligence, in line with organizations' security and autonomy needs.