Study shows potential for using AI tools to detect health care-associated infections

A brand new proof-of-concept study revealed within the American Journal of An infection Management experiences that synthetic intelligence (AI) applied sciences can precisely determine instances of well being care-associated infections (HAI) even in complicated scientific situations. The research, which highlights the necessity for clear and constant language when utilizing AI instruments for this function, illustrates the potential for incorporating AI expertise as a cheap element of routine an infection surveillance packages.

In response to the newest HAI Hospital Prevalence Survey carried out by the Facilities for Illness Management and Prevention, there have been roughly 687,000 HAIs in acute care hospitals within the U.S. and 72,000 HAI-related deaths amongst hospital sufferers in 2015. About 3% of all hospital patients have at the very least one HAI at any given time.

The implementation of infection surveillance packages and different infection-prevention protocols has diminished the incidence of HAIs, however they continue to be a danger, notably to critically in poor health hospitalized sufferers with inserted units akin to central traces, catheters, or respiratory tubes.

Many hospitals and different health care facilities have HAI surveillance packages to watch for elevated an infection danger, however they require intensive assets, coaching, and experience to keep up. In resource-constrained settings, a cheap various might assist to boost surveillance packages and permit for higher safety of high-risk sufferers.

On this new research, researchers at Saint Louis College and the College of Louisville College of Drugs evaluated the efficiency of two AI-powered instruments for correct identification of HAIs. One software was constructed utilizing OpenAI’s ChatGPT Plus and the opposite was developed utilizing an open-source giant language mannequin often known as Mixtral 8x7B.

The instruments have been examined on two varieties of HAIs: central line-associated bloodstream an infection (CLABSI) and catheter-associated urinary tract an infection (CAUTI). Descriptions of six fictional affected person situations with various ranges of complexity have been offered to the AI instruments, which have been then requested whether or not the descriptions represented a CLABSI or a CAUTI. The descriptions included info such because the affected person’s age, signs, date of admission, and dates that central traces or catheters have been inserted and eliminated. AI responses have been in comparison with professional solutions to find out accuracy.

For all six instances, each AI instruments precisely recognized the HAI when given clear prompts. Importantly, the researchers discovered that lacking or ambiguous info within the descriptions might forestall the AI instruments from producing correct outcomes. For instance, one description didn’t embody the date a catheter was inserted; with out that element the AI software couldn’t give an accurate response. Abbreviations, lack of specificity, use of particular characters, and dates reported in numeric format as an alternative of with the month spelled out all led to inconsistent responses.

“Our outcomes are the primary to exhibit the facility of AI-assisted HAI surveillance within the well being care setting, however in addition they underscore the necessity for human oversight of this expertise,” mentioned Timothy L. Wiemken, Ph.D., MPH, an affiliate professor within the division of infectious ailments, allergy, and immunology at Saint Louis College and lead writer of the paper. “With the speedy evolution of the function of AI in medication, our proof-of-concept research validates the necessity for continued improvement of AI instruments with real-world affected person information to assist an infection preventionists.”

Further particulars in regards to the research embody:

• Each AI instruments have been used with retrieval augmented era, an strategy that improves the standard of prompting via a information repository that offers the AI software further context. On this case, the repository included materials from CDC’s Nationwide Well being care Security Community, a monitoring system for HAIs.
• The ChatGPT Plus software developed for this research, HAI Help, is obtainable on the OpenAI GPT Retailer for these with a ChatGPT Plus subscription.

“HAI surveillance is a important accountability for an infection preventionists, and our group wants each doable software to assist us guarantee the security of our sufferers,” mentioned Tania Bubb, Ph.D., RN, CIC, FAPIC, 2024 APIC president. “This research means that AI-powered instruments could provide a cheap technique of bettering our surveillance packages by helping an infection preventionists in day-to-day work features.”

Offered by
Affiliation for Professionals in An infection Management