I don ’ t think it is complete enough at this point to rely solely on technology , but every journey has a starting point . Having IPs involved in developing processes to both train and perform / assist with performance of surveillance activities could be a tremendous benefit for the current workload that confronts typical IP & C practice .”
— Ruth Carrico , PhD , DNP at the OpenAI GPT Store for people with a ChatGPT Plus subscription .
Carrico emphasizes that AI can be a big asset to surveillance tasks undertaken by IPs which constitute a significant portion of their daily workload .
“ There is an ability to infuse more standardization into the use of HAI definitions , and that is often a challenge for IPs ,” she says . “ Many of us ( IPs ) look at the clinical condition and presentation of the individual patient and some of those factors may influence whether or not we consider an event to be an HAI . For example , we may have a patient with a common skin organism in multiple blood cultures . We may have concerns as to the conditions under which those cultures were obtained and that concern may lead us to disregard that particular HAI , even if it meets the case definition . The standardization that can be reinforced through the use of AI elements can help new IPs become comfortable with the purposes of surveillance and therefore less likely to adjudicate findings . Also , this is a great teaching point when IPs get together with others within our facility to review surveillance and surveillance processes and outcomes . It may be easier to have the discussion with a surgeon when reviewing SSI surveillance results . We are looking for risk factors we can address and not pointing fingers . Perhaps this approach can take a step forward in deeper collaboration .”
But is AI reliable enough so that it can be used to address the workload of IPs ?
“ I don ’ t think it is complete enough at this point to rely solely on technology , but every journey has a starting point ,” Carrico says . “ Having IPs involved in developing processes to both train and perform / assist with performance of surveillance activities could be a tremendous benefit for the current workload that confronts typical IP & C practice . This can be helpful across all settings and may provide new insights into processes and outcomes in settings where surveillance has been less actively or consistently performed .”
Wiemken says he agrees with Carrico . “ There is still a lot we don ’ t understand about the underpinnings of even some of the large language models . When moving to more fully generative AI approaches outside of text to text , there is even more to learn before we can start being confident enough to ‘ trust ’ the models with various
16 • www . healthcarehygienemagazine . com • june 2024
critical aspects of our workflows . There are certainly many limitations to any AI currently available — not the least of which is the seemingly constant shifts in capabilities . It can be hard to develop something that will have any true longevity - particularly since accuracy is very important . Massive gains in many areas of AI make it nearly impossible to develop something that will be seen as ‘ current ’ even in the next six months . Because of that , I think incremental improvements and proof of concept studies like what we conducted will be the most reasonable approaches in the near term .”
Balancing human know-how and technology will be the challenge for the future , the researchers say .
“ This is a good opportunity to pause and ask what support is available to the IP ,” Carrico adds . “ Right now , nothing takes the place of the collaboration and practice assessment that IPs lead , but how this is done in the future will undoubtedly change as we learn more about how to use what we are quickly learning in this field . It is a time for us to get rid of the old vision of ‘ Terminator ’ technology and spend time learning about the AI field and reimagine how our jobs can change . In my career , I have long recognized that things change . My goal now is to have change occur with design and not by accident .”
Wiemken says he is not concerned about technology taking over and doesn ’ t believe that AI could become a crutch . “ It is more of a retrospective confirmation bias — having information at our fingertips has likely made everyone more effective . It shifts the approaches we take to answering the critical questions we have . AI will serve the same purpose . It will shift workflows quite dramatically over time but that isn ’ t necessarily a bad thing . It should allow us to focus on other aspects of our work that either arise as new issues or things we have not had enough time to focus on historically because of bottlenecks .”
The researchers emphasize the importance of ensuring that AI-directed queries are clear and specific , to garner the most accurate results .
“ Clarity is important and that means we must ‘ learn the rules ’ of AI so we can apply them ,” Carrico says . “ We learned so much from this initial research , particularly in terms of the basics of developing and training these AI agents so they operate
as expected — with accuracy . This is a time of experimentation and learning , and that is the most exciting time .”
Wiemken adds that these issues will be reduced over time as things become more standardized and models converge into some common structure . “ This applies to how we interact with AI , but not necessarily the AI itself ,” he says . “ The ‘ garbage-ingarbage-out ’ saying will have more to do with what foundational model is being used . How was that model trained ? What data went into it ? How were biases addressed ? How high quality were the inputs ? All these things are critical to ensure you don ’ t have a highly biased or discriminatory model up front - at that point , what you put into it doesn ’ t matter , because you will get garbage out regardless . If you don ’ t understand the base model up front , the output can be hard to interpret even if it looks legitimate .”
Some experts are debating whether AI could add to the pressures of under-resourcing / training within the IP & C sector / general healthcare or relieve these burdens .
“ Everyone is always looking for ways to control and reduce costs ,” Carrico points out . “ IPs need to learn to be as budget savvy as possible . Get to know others within the facility who can help look at costs . For example , value analysis professionals can be a great partner with understanding supply chain disruptions . Modeling can help project product use and surveillance findings may bring additional information that is necessary for these projections . For example , the use of AI in helping frontline staff know what types of PPE are needed , and when , plays into the supply chain process . Once you begin to think about how AI can support our work through new ways of evaluating information , a whole new world for IPs seems to be opening . It is an exciting time to be in the field .”
Wiemken agrees , noting , “ We are just breaking ground on all of this , so the sky is the limit with application . Even outside of language models , image , audio , video , and other generative models can revolutionize nearly every area of our work .”
Other researchers agree on the contributions that AI can make toward improved surveillance efforts , which include “ interpreting databases generated from multiple data sources to prospectively monitor trends , identifying clusters and outbreaks in a timely fashion , tracking the
Continued on Page 18