Large language models for healthcare-associated infections (HAI) management: current applications, performance, and implementation evidence

Healthcare-associated infections (HAI) remain a major challenge to patient safety, requiring efficient synthesis of complex clinical data. Large language models (LLMs) have shown promise in medical applications, yet their role in HAI prevention and control is not well defined. To systematically identify and describe the applications, performance, and implementation characteristics of LLMs in HAI-related tasks. A systematic review was conducted following PRISMA guidelines and registered in PROSPERO (CRD420251031613). PubMed, Embase, Scopus, Web of Science, and IEEE Xplore were searched for studies published between January 2020 and February 2026. Study screening and data extraction were performed independently by two reviewers. Findings were synthesized qualitatively, and methodological reporting was assessed using an author-developed checklist informed by clinical AI reporting guidance. Fourteen studies were included, covering knowledge retrieval and question answering, LLM-assisted HAI surveillance, and clinical decision support. LLMs demonstrated generally high accuracy in guideline-based question-answering tasks (approximately 83% accuracy and a mean accuracy rating of approximately 4.0/5). In HAI surveillance tasks, sensitivity ranged from 80% to 100% and specificity from 35% to 86%, while review time per case decreased from 25 to 75 min with manual review to 5–7 min with LLM assistance. Accuracy was as low as 45% in complex tasks requiring temporal reasoning. All studies clearly reported the target task; however, none fully reported validation and safety assessments or evaluated prospective real-world clinical implementation. LLMs show strong potential in structured and screening tasks but remain limited in complex clinical reasoning. A task-complexity–based taxonomy of HAI-related LLM applications suggests their current role as assistive tools rather than autonomous decision-makers, underscoring the need for real-world validation and human–AI collaboration.