The Critical Role of Model Selection in Causal Inference: A Comparative Analysis of Classification Models within the InferBERT Framework for Pharmacovigilance
A new study published on arXiv evaluates the effectiveness of different classification models within the InferBERT framework for identifying causal adverse drug events (ADEs) in pharmacovigilance. The research found that domain-specific pre-trained models, such as BioBERT, significantly outperformed larger, general-purpose LLMs like Med-LLaMA and simpler models like XGBoost in accuracy and concordance with traditional pharmacovigilance signals. The study also indicated that while post-hoc calibration can improve model calibration, it has mixed effects on accuracy and causal discovery, suggesting that investing in domain-aware models is more beneficial than simply increasing model size for this application. AI
IMPACT Domain-specific pre-trained models show superior performance over general LLMs for causal inference in pharmacovigilance, guiding future model selection for specialized AI applications.