Mitigating hallucinations in healthcare AI: a systematic review of evidence-based strategies

Abstract Background Large language models (LLMs) are increasingly integrated into healthcare applications, but their tendency to generate hallucinations—factually incorrect yet plausible outputs—poses risks to safety and trust. In clinical settings, these failures include fabricated citations, incorrect treatment statements, and inaccurate summaries of patient context, each of which can propagate unsafe decisions. Methods Using the PRISMA systematic review guidelines, we reviewed empirical studies from January 2019 to April 2025 that evaluated hallucination mitigation strategies in LLM-based healthcare AI systems. We searched PubMed, IEEE Xplore, ACM Digital Library, Scopus, Web of Science, and arXiv using predefined terms. Eligible studies were assessed for methodological quality using an adapted Joanna Briggs Institute checklist. Results Out of 427 retrieved studies, 44 met inclusion criteria. We identified seven main strategy categories: (1) retrieval-augmented generation (RAG), (2) knowledge graph integration, (3) self-reflection frameworks, (4) specialized evaluation metrics, (5) human-in-the-loop (HITL) approaches, (6) specialized training techniques, and (7) red teaming. The most frequently evaluated strategies were RAG (18/44), knowledge graphs (12/44), and self-reflection or specialized training (10/44 each). RAG approaches demonstrated 30–50% reductions in hallucinations; HITL strategies achieved up to 95% reductions but posed scalability concerns; and red teaming identified novel vulnerabilities that other approaches missed, with interdisciplinary teams showing 20–40% improvements in hallucination detection. Combined approaches (e.g. retrieval plus verification) generally outperformed single-method controls. Conclusions Mitigating hallucinations in healthcare AI requires integrated approaches that combine technical and human-centered safeguards. We propose a typology of mitigation strategies, implementation considerations, and a replicable workflow to inform safer deployment in clinical settings.