A computationally efficient biomedical text processing framework for pharmacovigilance: integrating low-rank adaptation and interpretable AI for adverse drug reaction detection

Abstract Early detection of adverse drug reactions (ADRs) is crucial for patient safety but remains challenging due to underreporting and delayed data in traditional pharmacovigilance. This study proposes a computationally efficient and interpretable framework for ADR detection by integrating Low-Rank Adaptation (LoRA) and SHapley Additive Explanations (SHAP) with encoder-based transformer models (BERT, DistilBERT, RoBERTa). Leveraging over 3,900 annotated tweets, our approach demonstrates that LoRA reduces trainable parameters and training costs by up to 50%, while preserving high classification accuracy (above 98%) across three disease classes. SHAP analysis provides actionable interpretability, revealing that the models consistently rely on clinically relevant terms, such as drug names and symptoms, to drive predictions. Compared to traditional finetuning, LoRA and Efficient Finetuning of Quantized LLMs (QLoRA) offer a robust and scalable alternative for processing noisy, informal social media data, making real-time ADR monitoring feasible in resource-constrained healthcare settings. This framework strikes a balance between computational efficiency, interpretability, and predictive performance, supporting the integration of pharmacovigilance into clinical decision support systems for safer patient care. Graphical Abstract