Author: Chang Ho Yoon
Date: 20th March 2025

Large language models are poised to become valuable allies in infection prevention and control (IPC). The jury will be out for some time but the evidence to date is already hinting at an enticing, brave new world of increased efficiency (think real-time screening for HAIs and auto-generating IPC reports) and improved detection of cases otherwise missed or delayed by traditional – i.e., manual – methods​ (Rodriguez-Nava et al. 2024; Boussina et al. 2024).

What’s the main innovation here? The analysis of complex unstructured and structured data often centred around language. At a high-level, the basic idea is to serve as on-demand knowledge resources for busy IPC team members. When compared to existing tools, LLM-based approaches often match or surpass performance benchmarks, for instance, an LLM system matching expert reviewers in 90% of sepsis cases (Boussina et al. 2024), or an AI model beating standard early warning scores in predicting infections​ (Creagh et al. 2024). If we free clinicians from tedious documentation and digging through stodgy interfaces to find the data they need, can they provide quicker, more comprehensive IPC needs? Can we predict and prevent outbreaks within hospitals before they happen?

“for instance, an LLM system [matches] expert reviewers in 90% of sepsis cases” - Aaron Boussina

LLMs are looking especially promising, but definitive empirical validation isn’t out there yet. Additionally, enthusiasm must be tempered by realism about current limitations. LLMs are not magic bullets – they sometimes fail in the exact scenarios where IPC cannot afford errors (complex, high-stakes decisions). The literature urges an augmented approach: use LLMs to handle workload and crunch data, but keep humans in the loop for oversight, validation, and the nuanced decision-making that machines can’t replicate (Wiemken & Carrico 2024; Schwartz et al. 2024).

With this approach, the IPC community can harness LLMs as “co-pilots” in infection control: the AI handles the mundane and provides preliminary analyses, while IPC teams remain the ultimate decision-makers ensuring patient safety. Going forward, we clearly can’t rest on our laurels – continued research, real-world evaluations, and interdisciplinary collaboration (between, e.g., AI developers, IPC experts, policy-makers, ethicists) will be crucial.

LLMs have arrived and, even if they are soon to be subsumed by more general artificial intelligence models, the linguistic component will remain front and centre. At every turn, however, it’s our prerogative to step back and question the co-pilot status of these “knowledge companions”: would you trust your second-in-command with the lives of so many?