Language models in pharmacovigilance: Applications, promises and limits

The integration of language models into pharmacovigilance offers a valuable opportunity to enhance quality and efficiency across workflows. With their rapid evolution, these techniques have found diverse applications in pharmacovigilance, revealing promising advances in time-consuming and low-added-value tasks; nevertheless, several practical constraints continue to hinder their adoption. This work examines the use of traditional natural language processing techniques and advanced language models across three major pharmacovigilance domains: (i) adverse drug reactions extraction from medical records, (ii) case processing, and (iii) evidence screening. A PubMed® search was conducted to identify potentially relevant studies. Subsequently, expert-based selection refined the core literature set, which was expanded through related-reference screening to capture highly cited or conceptually linked papers. Traditional natural language processing techniques, including rule-based systems, dictionaries, and statistical models, offer transparent and efficient mechanisms for structured information extraction, and have shown practical applications, particularly in adverse drug reaction identification and coding. However, they often struggle with the linguistic variability typical of clinical narratives. In contrast, advanced language models, including large pre-trained transformers and large language models, demonstrate superior contextual understanding and adaptability to unstructured and heterogeneous text sources. Yet, their regulatory acceptance remains limited by hallucination risks, reduced transparency/reproducibility, dependence on parameter tuning, and the continued need for human-in-the-loop validation to ensure reliability. Additionally, the substantial computational requirements of large language models impose significant environmental costs, emphasizing the importance of rational and sustainable implementation strategies. Overall, natural language processing and advanced language models should be regarded as complementary approaches. Their integration can augment human expertise and foster a scalable, trustworthy, and sustainable pharmacovigilance ecosystem. At present, an interdisciplinary approach, where pharmacovigilance professionals actively contribute to model design, validation, and oversight, remains essential to harness automation benefits while maintaining clinical integrity.