Large Language Models Show Promise and Limits in Hospital Infection Control

Large language models are increasingly being tested against one of hospitals’ most persistent safety problems: healthcare-associated infections, or HAIs. A systematic review published in Antimicrobial Resistance & Infection Control has pulled together the available evidence on how these AI systems perform in infection prevention and control tasks, and the picture that emerges is one of genuine promise tempered by substantial limitations. The review, registered in PROSPERO and conducted according to PRISMA guidelines, found fourteen studies published between January 2020 and February 2026 that applied LLMs to HAI-related work.

The included studies clustered around three main application areas. The first was knowledge retrieval and question answering, where models were queried about infection control guidelines and best practices. The second was LLM-assisted surveillance, in which models helped identify infections from clinical data. The third was clinical decision support, where the tools were positioned to aid clinicians managing patients with or at risk of healthcare-associated infections. This grouping allowed the authors to build a taxonomy organized around task complexity, an approach that shapes their central conclusion about where these models currently belong in clinical workflows.

Performance varied sharply with the difficulty of the task. In guideline-based question-answering, the models performed well, achieving approximately 83 percent accuracy and a mean accuracy rating of roughly 4.0 on a 5-point scale. For infection control teams needing rapid answers about isolation precautions or guideline recommendations, the reviewed evidence suggests LLMs can retrieve and relay such information with reasonable reliability.

Surveillance results were more mixed. Across the included studies, sensitivity in HAI surveillance tasks ranged from 80 to 100 percent, meaning the models generally caught most infections they were meant to flag. Specificity, however, ranged far more widely, from 35 to 86 percent.

One of the most concrete findings concerned workload. Manual review of cases in surveillance took between 25 and 75 minutes per case in the reviewed studies. With LLM assistance, that time dropped to 5 to 7 minutes per case.

Yet the review also documented a hard ceiling. In complex tasks requiring temporal reasoning, such as piecing together the sequence of events in a patient’s record to determine whether an infection meets surveillance definitions, accuracy fell as low as 45 percent. That is far below what clinical use would demand, and it illustrates a known weakness of language models: reasoning over time-ordered clinical events, where subtle details of timing and documentation determine the correct answer, remains unreliable.

The methodological assessment added another layer of caution. Using a checklist the authors developed based on clinical AI reporting guidance, they found that all fourteen studies clearly reported the target task, which is a basic but essential element of transparency. However, none of the studies fully reported validation and safety assessments, and none evaluated prospective real-world clinical implementation. In other words, every result in the review comes from retrospective or controlled evaluation contexts, not from live deployment in hospitals where errors would affect actual patients and workflows.

The authors’ overall conclusion is measured. LLMs show strong potential in structured and screening tasks but remain limited in complex clinical reasoning. Their task-complexity-based taxonomy supports framing these systems as assistive tools rather than autonomous decision-makers, with humans retaining responsibility for final judgments. The review underscores the need for real-world validation studies and for designing workflows around human-AI collaboration rather than replacement.

For infection prevention and control programs watching the rapid advance of medical AI, the practical takeaway is to treat current LLM performance figures as early and conditional. The strongest evidence supports using these models to accelerate guideline question answering and to speed the initial screening steps of surveillance, while complex case adjudication should stay with trained human reviewers until prospective studies demonstrate otherwise. As the field moves from proof-of-concept studies toward deployment, the review provides a baseline: fourteen studies, clear reporting of tasks, and a conspicuous absence of the safety and implementation evidence that real clinical adoption will require.

Subject of Research: Applications, performance, and implementation evidence of large language models in healthcare-associated infection management

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

Article References: Shan, J., Wang, H., Bao, X., Jin, M., Huai, W., Jin, Y., Ren, Y., Li, H., & Cao, Y. (2026). Large language models for healthcare-associated infections (HAI) management: current applications, performance, and implementation evidence. Antimicrobial Resistance & Infection Control. https://doi.org/10.1186/s13756-026-01828-2

Image Credits: AI Generated

DOI: 10.1186/s13756-026-01828-2

Keywords: large language models, healthcare-associated infections, infection control, HAI surveillance, clinical decision support, systematic review, artificial intelligence, patient safety, hospital epidemiology, AI validation, Large, language