TOWARDS SAFETY AND TRUSTWORTHY IN BIOMEDICAL AI

Abstract

Large Language Models (LLMs) have emerged as powerful tools in the medical domain, demonstrating human-level capabilities that enable applications ranging from clinical trial matching to risk prediction, biomedical knowledge retrieval and gene reasoning. These advances position LLMs as promising agents in healthcare, yet their safe and trustworthy deployment remains hindered by critical challenges, including bias, robustness, and patient privacy.This dissertation investigates the limitations of LLMs, including fairness and adversarial manipulation, in clinical contexts through both empirical analysis and system design. To systematically evaluate AI safety concerns, we propose a framework for trustworthy medical AI grounded in five core principles: Truthfulness, Resilience, Fairness, Robustness, and Privacy. Within this framework, we introduce a comprehensive benchmark of 1,000 expert-verified clinical questions designed to assess model behavior under sensitive scenarios. Finally, as a step towards addressing safety problems, we propose a novel inference-time method that significantly reduces memorization risk while preserving medical task performance, without the need for model retraining. Together, these contributions establish a computational foundation for evaluating and improving the safety of biomedical LLMs. The work advances the development of systems that are not only high-performing but also equitable, privacy-conscious, and aligned with the ethical and regulatory standards required for responsible clinical deployment.