The Impact of Generative Artificial Intelligence on Health Care

By Susan Chapman, MA, MFA, PGYT for For The Record

Artificial intelligence (AI) is an evolving technology that’s increasingly being employed across the health care industry. Like many technologies, AI continues to evolve, moving from what’s commonly referred to as traditional AI to its next generation, Generative (Gen) AI. When applied correctly, Gen AI can positively influence revenue cycle efficiency, clinical support, patient engagement and communication, and other areas of health care.

Traditional AI and Generative AI
Traditional AI has been in place in health care for about the past 10 years. The technology was designed around prediction and decision-making tasks, relying on structured, organized data that’s rules-based across a set of specific parameters. “As an example, if you wanted to predict if somebody could reappear in the hospital setting, you could build a predictive AI model to define that,” according to Liam Bouchier, Impact Advisors’ managing director of data and AI. “More generally, some basic AI that most people are familiar with are things like a spam filter on email or your phone. It’s machine learning that gradually learns every time someone moves something to the spam folder or blocks a call. It gradually improves, and some filters are better than others, which means it’s not something we can completely trust. So, traditional AI really focuses on one action, and a person has always been involved in influencing that action very early on in the process.”

Gen AI creates, or generates, new content and does not require structured information to train models. It can analyze a variety of data that is structured or unstructured—for example, audio images, notes, or books. From that information, it can decipher patterns and continue to learn.

Bouchier recently coauthored a white paper published by his company that summarizes Gen AI and its potential to transform the health care field. He and his coauthors wrote, “Gen AI represents a significant advancement in the field of [AI] and fundamentally changes how business processes are designed, executed, and optimized. While machine learning, natural language processing, and computer vision have been used in healthcare [sic] over the past decade, those instances have been limited and, in many cases, have represented expensive investments that did not always result in lasting value or transformative change for many early adopters.”1

But Gen AI is different. “We’ve already seen some of its models in general use with organizations that have ingested all of the published information available from the internet to train their models,” Bouchier says. “That’s the exciting part of it. Another main difference between traditional and Gen AI, is that the latter uses complex algorithms designed to mimic human behavior. It’s got something called neural networks, which is the structure behind all of the algorithms that are deployed across the information.”

In the past in the revenue-cycle world, key performance indicators relied on human ability to complete the coding process on tight deadlines, using traditional AI. “However, when an account can be coded entirely autonomously, this removes the need for a human touch, and propels a quicker coding-to-billed times for an organization,” notes Diana Ortiz, RN, JD, CCDS, CCDS-O, senior manager of global content at Solventum, formerly 3M Health Care. “For those accounts that don’t meet the thresholds set for autonomous standards, there are still a lot of efficiencies that can be gained by having the majority of codes assigned autonomously and leave the human touch to more efficient final quality reviews, to obtain a more specific code than presented, or resolve a coding/billing edit. The accuracy around AI in the coding space comes from leveraging massive amounts of data to train models to determine accurate and valid codes that should be included on a claim. Bringing together clinical, coding, and payer insights is key to leveraging multiple models at scale to determine high levels of accuracy, as has long been demanded in the HIM industry.”

Claims management and denial prevention are two areas that can benefit from Gen AI. “One real-world example is a health system in Chicago where a Gen AI model was deployed. That system experienced a 5% reduction in denials for that system’s top five denial codes and recouped up to 14 million in potential lost revenue,” Bouchier says. “What’s especially interesting is that the only thing the model was doing was recognizing what those codes were and identifying what the missing documentation was. It then analyzed the chart, gathered the correct information, and sent it to the payer. The payer didn’t deny the claims because the supporting documentation, which had been absent in the past, was there.”

In clinical support, traditional AI has already been prevalent in many areas, but especially in radiology. Traditional AI can analyze an image and provide a decision based on the information available at that time. Gen AI, though, is beginning not only to look and make predictions based on the current image but also to guide radiologists through the image. “It can tell the radiologist how to look at the image and why to analyze the image in a certain way. The technology would do that based on the history of radiologists discovering certain elements,” Bouchier says. “That is a good example of how Gen AI is helping providers move to the next level of decision-making based on what it has seen in the past. It’s not just giving clinicians information from a particular point in time.”

Patient care and engagement extend beyond the revenue cycle, and health care systems are seeking ways to simplify their communications and enhance their engagement with their patient populations, not just for collections. Therefore, another area where health systems are seeking to deploy this advanced technology is with patient communication, for instance, in simplifying things like explanation of benefits, which have become increasingly complex for patients to understand. Gen AI can drive meaningful value in streamlining those communications touch points; creating personalized documents for patients, providers, and payers; generating real-time treatment plans and summaries; and assisting with onboarding and postdischarge tasks, including patient medication adherence. Additionally, Gen AI can play a role in virtual medicine, offering opportunities to provide 24/7 support to health care professionals through the automation of routine tasks, which can then increase productivity and improve patient care.1

“Hospitals also want to prevent readmissions, and Gen AI can play a significant role in those types of things,” Bouchier states. “For instance, when Gen AI replaced a previous AI model in one health care system, the preventable readmission rate was improved by 5%, which is statistically an enormous improvement. The cost associated with that is huge, given that in the United States, the average readmission costs about $15,000. Health care systems are penalized for preventable readmissions, and this technology can have a significant impact very quickly on how systems are being measured by CMS. There are some good opportunities on the horizon in many areas.”

Accessibility for Large and Small Organizations
Much of this technology is available as open source and deployable on an organization’s existing cloud architecture. Health care organizations only then need data to train the system. Therefore, according to Ortiz, this emerging technology should be available to organizations of any size. “What it really comes down to is having enough volume to train an AI model. So instead of thinking initially about leveraging this type of technology broadly within an organization, the approach should be largely driven by service lines where automation would generate a large return on investment [ROI],” she says. “For example, the volume of emergency-department visits that need to be billed for both professional and facility coding may require a large number of FTEs [fulltime equivalent employees] to support the effort, and could be an area of automation that would drive positive results and efficiencies. Whereas a small-clinic setting with a couple of providers conducting routine office visits, may require a lot of time and volume to train a model, and yield very little ROI. So it really isn’t about facility size, but more driven by service line volumes as well as perspective around ROI.”

Potential for Both Bias and Progress
While Gen AI offers a great deal of promise in the health care field because it mimics human behavior, it’s at great risk for bias depending on a number of different factors. “That’s one of the major concerns in the health care market,” Bouchier says. “So responsible AI is becoming a common term in the field. We’ve built this tool, but how do we know that if we apply it to a certain patient population it’s going to behave in the way we hope is not biased? That will depend on a number of things, especially the information used to train the model.”

For many reasons, including the potential for bias, the information that’s used to train the technology is the most fundamental and significant aspect of Gen AI. But because Gen AI has the ability to train on data specific to a health care system’s population, the technology is gaining traction across the industry. “Let’s say I live in a particular part of the country that has a particular socioeconomic demographic that is different from another part of the country, Los Angeles [L.A.] vs New Orleans, or New York, L.A., and New Orleans. They are very different populations with different demographics, and the care provided in those areas may be standard care but there are nuances regarding how that care is delivered,” Bouchier explains. “It’s exciting for health systems because they can analyze their unique populations to provide a better, more customized set of services.

Another focus for Gen AI could be in the quality arena. Adverse quality events occur in many health care organizations and there are a subseries of quality metrics that hospitals track and have to report on to CMS on a regular basis. To improve processes and prevent the reoccurrence of one of those events, a health system will perform an audit to determine what happened. “For example, patient falls. If a patient fall happens, the organization will perform an audit to try to understand what happened, which is a fairly lengthy, labor-intensive process,” Bouchier says. “Even though the records are all there electronically, you’ve got to put guardrails around it and figure out the people [care teams], procedures, and policies involved. Gen AI can look at all of those vast amounts of inputs and data behind that incident because the volume or structure of data is not a limiting factor. All of that information can be used to train the large language model. You can then ask it questions, like if the policies and procedures were up-to-date and were they appropriately followed. The system can deliver accurate answers instantly, which makes the technology an immense time-saver for these audits.”

Gen AI’s Current and Future Impacts
Given Gen AI’s relatively new place in the health care industry, experts believe there’s not necessarily enough information to measure its impact accurately. Bouchier notes that he has seen estimates of up to a 50% reduction in preventable readmission rates in the United States. “It’s all very new, and there are many different figures out there,” he says. “There is still going to be an adoption piece. A recent Pew Research poll showed that 60% of Americans are uncomfortable with health care providers relying on AI. Thirty-eight percent of Americans believe that AI will lead to better patient outcomes. We’re only about 12 months into probably a five-to-10-year cycle. So, I don’t think there’s enough information yet. Maybe by year three or four of that cycle, we’ll be able to tell if it’s having an impact, at least from the administrative side of things. With the revenue cycle, though, that should be apparent faster because those are annual cycles. I think we could probably do that within 18 to 36 months.”

Bouchier notes that patient outcomes are different because they are complicated far beyond what one can see in just the health care environment, citing statistics that about 10% of a person’s experience with the health care system affects their health outcomes, with the remainder affected by other factors. “There are a lot of things that the health care system has no control over that can affect people, so I think that’s a much longer cycle,” he says. “However, with things like patient communications and making those aspects of health care easier to understand, better engagement for patients so people come back in for their follow-ups, and other things like that, we’ll see the impacts fairly quickly.”

Ortiz agrees that since the industry is still in the early stages of adopting this new technology, it may be difficult to assess its impact. “As more organizations are choosing whom to partner with on their journey toward automation, we will start to see the widespread impact in the coming years,” she adds. “As AI is applied successfully in the coding and HIM space, we will likely see expansion into many of the administrative tasks associated with the broad revenue cycle.

“There is also tremendous opportunity to apply AI in the payer-denials area, where much of the work is completed manually today and is incredibly expensive,” Ortiz continues. “However, it is a necessary task to collect accurate reimbursement for services provided. Many case-management and utilization management activities could benefit from more automated workflows, to create efficiencies for both patients and organizations looking to ensure proper level of care placement and correlating reimbursement. There are tremendous opportunities to ultimately provide better and more accurate clinical documentation and coding through the application of responsible AI.”