Leveraging Large Language Models for Adverse Drug Event Detection: A Comparative Study of Token and Span-Based Named Entity Recognition

Adverse Drug Events (ADEs) pose a persistent threat to patient safety and public health. This study investigates the use of large language models (LLMs) fine-tuned for both token classification and span-based named entity recognition (NER) to improve ADE detection in clinical text. Using the 2018 n2c2 Track 2 dataset, we evaluate models under both predefined (gold label) and end-to-end settings. RoBERTa Large consistently outperforms other models, particularly in identifying ADEs, which remain more challenging due to their contextual ambiguity. Token-based models generally deliver stronger performance than span-based approaches, and ensemble methods, especially majority voting and XGBoost-based aggregation, further enhance end-to-end relation extraction by mitigating individual model weaknesses. These findings highlight the potential of fine-tuned LLMs, augmented by strategic ensembling, to advance clinical NLP pipelines and support safer, more personalized healthcare.