DART: A Structured Dataset of Regulatory Drug Documents in Italian for Clinical NLP

TL;DR

DART addresses the shortage of Italian regulatory NLP resources by building a scalable, reproducible pipeline that harvests Italian RPCs from the AIFA repository, segments them into standard regulatory sections, and generates structured fields plus clinical summaries with few-shot LLMs. The dataset comprises 16,029 documents and over 95 million tokens, with high coverage of core sections and a rich, domain-specific vocabulary, enabling robust downstream tasks such as DDI detection, knowledge graph construction, and regulatory model tuning. A highlighted contribution is an LLM-driven drug–drug interaction checker that leverages DART-derived summaries to achieve significantly higher recall than baselines, demonstrating the practical impact of grounding AI systems in regulatory texts. By releasing the code and a comprehensive, multilingual resource, the work supports AI-enabled pharmacovigilance, regulatory reasoning, and safer clinical decision-support in the Italian healthcare setting.

Abstract

The extraction of pharmacological knowledge from regulatory documents has become a key focus in biomedical natural language processing, with applications ranging from adverse event monitoring to AI-assisted clinical decision support. However, research in this field has predominantly relied on English-language corpora such as DrugBank, leaving a significant gap in resources tailored to other healthcare systems. To address this limitation, we introduce DART (Drug Annotation from Regulatory Texts), the first structured corpus of Italian Summaries of Product Characteristics derived from the official repository of the Italian Medicines Agency (AIFA). The dataset was built through a reproducible pipeline encompassing web-scale document retrieval, semantic segmentation of regulatory sections, and clinical summarization using a few-shot-tuned large language model with low-temperature decoding. DART provides structured information on key pharmacological domains such as indications, adverse drug reactions, and drug-drug interactions. To validate its utility, we implemented an LLM-based drug interaction checker that leverages the dataset to infer clinically meaningful interactions. Experimental results show that instruction-tuned LLMs can accurately infer potential interactions and their clinical implications when grounded in the structured textual fields of DART. We publicly release our code on GitHub: https://github.com/PRAISELab-PicusLab/DART.

Figures (4)

• Figure 1: End-to-end pipeline for the construction of the DART dataset. The diagram outlines the three main components of the workflow: automated retrieval of RCP documents from AIFA, text extraction and segmentation, and the final structuring and validation of the resulting dataset.
• Figure 2: Two-step RESTful API interaction for programmatic retrieval of RCP PDFs from AIFA's infrastructure.
• Figure 3: System pipeline for automatic detection of drug–drug interactions (DDIs). Each commercial drug, associated with its active ingredient $(F_1, F_2, ..., F_N)$ is linked to its respective Summary of RCPs. From each RCP, the system extracts clinically and pharmacologically relevant sections. These sections are then processed by the Summarizer component, it is an LLM which generates a structured summary (max. 450 words). The resulting summaries are fed into a LLM, which analyzes all drug pairs and classifies the interaction severity as absent, minor, moderate, or major with prompt engineering.
• Figure 4: An end-to-end example of drug–drug interaction detection using the DART framework. The system retrieves, summarizes, and compares documents to assess clinical relevance.