Study Shows How AI Supports HAI Prevention

Earlier this year, researchers at Saint Louis University and the University of Louisville School of Medicine published a proof-of-concept study in the American Journal of Infection Control. The study found that artificial intelligence (AI) technologies can accurately identify cases of healthcare-associated infections (HAI) using gold standard surveillance definitions, even in complex clinical scenarios.   

This is a promising breakthrough in the complicated task of surveilling and preventing these devastating and, unfortunately, all-too-common infections. According to the most recent HAI Hospital Prevalence Survey conducted by the U.S. Centers for Disease Control and Prevention (CDC), there were approximately 687,000 HAIs in acute care hospitals in the U.S. and 72,000 HAI-related deaths among hospital patients in 2015. Meanwhile, 2022 reports indicate that, on any given day, roughly 1-in-31 hospital patients are dealing with at least one HAI.  

At present, many hospitals and healthcare facilities rely on HAI surveillance programs to monitor infections, allowing prevention and control programs to develop preventative interventions. A critical first step is the use of a standardized approach to identify infection and understand risk factors for their occurrence. However, surveillance programs are often burdensome, requiring extensive resources, training, and expertise to achieve and maintain. In resource-constrained settings, and truly in all settings, a cost-effective alternative could help enhance surveillance programs, allowing for more time to be spent in development of interventions and better protecting patients. 

Enter Artificial Intelligence (AI).  

The results of the study suggest that AI-powered tools can help streamline infection prevention by automating some tasks. That would be a welcome leap forward for beleaguered staff, as the report states that, “Surveillance activities have been ranked as one of the program responsibilities that account for between one-quarter and one-half of the personnel work hours, most often by individuals responsible for the facility infection prevention and control programs.” 

This means, in theory, AI-powered tools could free up the infection preventionist from some of their most time-consuming, resource-heavy surveillance work.   

“Surveillance is a tough nut to crack,” explains the study’s co-author Ruth M. Carrico, Ph.D., DNP. “Many in the field possess a strong clinical background. That means they are used to looking at the patient to see if they clinically meet a particular situation. Surveillance is very different. Surveillance casts a wide net to identify risk factors that may not have been recognized before. This is where infection preventionists struggle. But to do a good job you need to do both.”   

Consistent training and an ability to improve the quality of surveillance and verify competence would be a welcomed program element powered by AI. 

In their study, Carrico, along with lead author Timothy L. Wiemken, Ph.D., MPH, worked with OpenAI’s ChatGPT Plus and an open-source large language model known as Mixtral 8x7B. The pair tested the ability of these large language models to perform HAI surveillance. Results, with a few caveats, were extremely promising.  

Both AI models accurately identified Central Line-Associated Bloodstream Infections (CLABSI) and Catheter-Associated Urinary Tract Infections (CAUTI) from six National Health Care Safety Network training scenarios. However, there were some identified challenges. For example, feeding the AI tools incomplete or ambiguous information could prevent them from producing accurate results.  

Still, the study suggests future research possibilities for other HAI pressure points, like surgical site infections, laboratory identification of infection, and ventilator-associated events. Carrico expands on this idea.  

“Take a piece of reusable medical equipment for example,” she says. “Right now, we say, ‘treat it according to the Spaulding Classification.’ That’s very broad and strategic. But could we use AI to get down into the weeds? Can we ask AI, ‘How can I manage this specific piece of equipment to make it safe for the patient?’ Or, ‘Can we gather information regarding medical device use and include it as a potential risk factor?’ That could feed new information for the AI agent and perhaps give us a better idea of the risks involved when we reuse medical devices, and decide how and when to best clean and disinfect.” 

Carrico foresees many more opportunities for AI-assisted HAI surveillance work. One area she finds intriguing involves the current supply crisis involving blood culture bottles. Right now, only two companies make the laboratory machine capable of performing the majority of blood culture testing performed in the U.S. Each machine requires the use of their own custom-made bottle. But it has become a national problem as one of those companies encountered a supply chain snag.  

In this case, almost instantly, the capacity to analyze organisms was impacted. A critical element of a response process involves how to ensure every blood culture bottle used can produce a usable result. This means that performance of the process to collect the blood culture must be as close to perfect as possible.     

“A contaminated blood culture can be catastrophic for patient diagnosis and microbial identification,” Carrico says. “AI could be of great use here. Can we use new AI capabilities that allow the use of text, in this case a blood culture collection procedure, to be translated into images or videos? This way, we could see if the actions we are putting into those written procedures can be used to perform that procedure in a more perfect way. AI could also help establish better algorithms to identify predictors for when we should draw blood cultures and where are errors or performance can be improved.”