While there has been significant progress in preventing many health care-associated infections, they still pose a serious risk to patients, hospitals, and clinics. Artificial intelligence (AI) may offer a solution that increases surveillance and allows infection prevention specialists to engage in more hands-on and educational work.
Guillermo Rodriguez Nava, MD, assistant professor of infectious diseases at the University of Colorado Anschutz School of Medicine, says research increasingly shows that large language models (LLMs), a type of AI that is designed to understand and analyze human text, are promising tools for processing the vast amounts of unstructured clinical data generated in health care that is often needed in infection prevention work.
LLMs have been evaluated for health care-associated infection surveillance, mostly central line-associated bloodstream infections, surgical site infections, and catheter-associated urinary tract infections, with pooled sensitivities exceeding 90% across various studies.
“It’s important to remember that these algorithms are not thinking machines. AI is not here to replace a doctor, and I don't think they ever will, but they will help us improve in our workflow in infection prevention,” says Rodriguez, who is also the associate medical director of infection prevention and control at Denver Health.
Making time for enhanced strategy
Currently for most hospitals utilizing electronic health record (EHRs), the system screens charts for specific criteria, then flags them for potential infection risk.
With bloodstream infections, for example, if a patient has a central line while in the hospital and the EHR shows a positive blood culture, that will be flagged. The infection prevention specialist gets an alert and then reviews the chart to see whether the patient meets the surveillance definition outlined by the Centers for Disease Control and Prevention (CDC) and the National Healthcare Safety Network.
Most hospitals have a team of infection prevention specialists that review charts when flagged. This can make up most of their work shift, as the CDC estimates that approximately one in 38 hospital patients has at least one health care-associated infection on any given day.
These specialists must also strategize how to prevent these infections for patient safety and for the hospital or clinic, which can face fiscal or regulatory penalties if infection rates are deemed too high.
“In a lot of hospitals there’s limited time in the day to visit units and review external devices on the patient, develop new strategies to prevent infections, or educate staff about infection prevention strategies, like handwashing,” Rodriguez says.
Having fewer flagged EHRs to review, however, may allow specialists to do more of that.
Expanding progress
In the last several years, infection prevention in most places across the country has improved. The latest National and State Healthcare-Associated Infections Progress Report shows that acute care hospitals experienced an 11% decrease in C. difficle infection, 10% decrease in catheter-associated urinary tract infections, a 9% decrease in central line-associated bloodstream infections, and a 7% decrease in methicillin-resistant Staphylococcus aureus (MRSA) over the previous year.
Still, there’s room for progress, and because so much of infection prevention is based on strict definitions and criteria, AI can be a helpful tool, Rodriguez says. “We are not asking AI to make clinical decisions, but to summarize the data and to look for criteria and make a determination of whether it fits the surveillance definition outlined by the CDC.”
With the assistance of LLMs, this work can be done much faster, perhaps taking five minutes compared to the hour it would take a human infection prevention specialist to review without the assistance of an LLM. Rodriguez says that double-checking the work of AI tools is necessary just like it is in nearly all medical settings. Even so, evidence points to LLMs as being big timesavers.
“At the end of the day, you would only have to review the cases that the AI system marks as the true positives because they are very good at finding almost all true negatives,” Rodriguez says.
Validation and the future
EHRs are often comprised of unstructured clinical text — lots of notes that can span many days if the patient has a longer hospital stay. This text requires sorting and analysis to be beneficial for infection prevention. That’s a major benefit of using LLMs, because an AI system can more quickly maneuver through those notes, make sense of them, and determine whether the flagged EHR does meet standardized infection surveillance definitions.
Many hospital systems are already deploying their own AI tools, some reporting 100% sensitivity when it comes to detection. However, Rodriguez cautions that more validation is needed to ensure true accuracy and efficacy, especially because it can be difficult to replicate similar outcomes in different hospital settings.
“Before there’s a widespread adoption of AI in infection prevention, we should further validate, especially across diverse settings to ensure their use as a screening and decision-support tool,” Rodriguez says.
