Interpretable Biomedical QA: Customizing Models for PubMedQA with Knowledge Distillation and XAI

Abstract

Biomedical question answering (QA) is pivotal for advancing clinical decision-making, with PubMedQA serving as a critical benchmark to evaluate model performance in understanding complex medical texts. This paper presents a comprehensive study on customizing and evaluating state-of-the-art language models for PubMedQA, focusing on domainspecific pretraining and architectural optimizations. This study introduced novel experimental results from an extensive evaluation of models including BioMedBERT, PubMedBERT, SciBERT, BioBERT, and others, using the PubMedQA-Labeled (PQA-L) and PubMedQA-Artificial datasets. Our findings demonstrate significant performance improvements through techniques such as knowledge distillation, ensemble methods, Explainable AI and architectural modifications. The results underscore the potential of tailored models to democratize access to medical AI, particularly in resource-constrained settings. Leveraging Explainable AI (XAI), this research analyzed token-level attention weights to uncover critical factors influencing predictions, enhancing model interpretability. This approach challenges the black-box perception of AI, providing deeper insights into decision-making processes for biomedical QA.

Ähnliche Arbeiten