Overview of RAG-based and LLM-based approaches to personalization in healthcare AI applications

Abstract

Recent advances in Large Language Models (LLMs), driven by transformer architectures such as Generative Pre-Trained Transformer (GPT), are opening new frontiers in healthcare Artificial Intelligence (AI) by enabling clinically relevant interactions between patients and clinicians. Yet persistent challenges—including limited real-time knowledge access, safety concerns and insufficient patient-centered contextualization—indicate that current systems often fall short in delivering efficient and reliably personalized responses (Li et al., 2025; Cascella et al., 2023). To address these gaps, Retrieval-Augmented Generation (RAG) can couple LLMs with external knowledge sources at inference time, producing responses that are more grounded, up-to-date, and tailored to individual patient needs. This systematic review examines how personalization is operationalized in LLM-only and RAG-enhanced healthcare AI systems. We searched multiple scholarly sources and included 20 studies in the final qualitative synthesis. We compare personalization strategies, evaluation practices, and trustworthiness challenges, with emphasis on healthcare-specific issues such as privacy, safety, and clinically meaningful context integration. We analyze how current evaluation frameworks assess these systems and identify limitations in their ability to reflect clinically meaningful performance. We further include a brief MedQuAD-based illustrative case study to highlight limitations of current evaluation metrics, demonstrating that strong benchmark performance does not necessarily correspond to clinically reliable or deployment-ready behavior. The review finds that personalization remains inconsistently defined and weakly evaluated in healthcare AI, with most systems implementing retrieval-grounded adaptation rather than true patient-specific personalization. Furthermore, relatively few studies directly evaluate hallucination, patient safety, clinician validation, or deployment robustness, revealing a critical gap between commonly reported performance metrics and the requirements of clinically reliable, safe, and real-world deployable decision-support systems.