Generative Artificial Intelligence Applied to Patient Safety and the Detection of Adverse Drug Events

Abstract

Patient safety is one of the major challenges facing contemporary healthcare systems, requiring strategies capable of the early identification of risks and adverse events during healthcare delivery. Although electronic health records (EHRs) contain a large volume of clinical information that could support this purpose, manual review of these records is time-consuming, costly, and impractical on a large scale, limiting their use for surveillance and clinical decision-making. In this context, Large Language Models (LLMs) have emerged as one of the most promising applications of artificial intelligence for the automated interpretation of clinical documentation. However, most currently available models have been developed for other languages and healthcare settings, and there are currently no validated technologies for Brazilian Portuguese capable of automatically identifying adverse drug events (ADEs) with sufficient performance to support patient safety programs. This gap highlights the need to develop models specifically adapted to the Brazilian healthcare context. The hypothesis of this study is that fine-tuning a Brazilian clinical language model, combined with diagnostic validation by clinical experts, will improve the accuracy of automated ADE identification. This hypothesis is supported by recent evidence from the literature and by preliminary findings from our research group, which demonstrated the feasibility of using a LLaMA-based model for ADE screening in pediatric electronic health records, achieving high sensitivity for event detection but limited precision, indicating the need for further refinement through fine-tuning. To test this hypothesis, a methodological study on the development and diagnostic validation of a health technology will be conducted in three complementary phases, following the STARD recommendations. The expected outcome is to improve ADE detection and deliver a validated national technology with the potential to be incorporated into patient safety programs.