Generative AI’s Potential Shines on Revenue Cycle Management

Fact or fiction?

According to ChatGPT, a popular generative AI tool right now, generative AI is “a class of artificial intelligence techniques that involve training models to produce novel data samples or content, often in the form of text, images, music, or other media types.”

“It leverages deep learning algorithms, particularly variants of neural networks, to generate content that is not directly copied from existing examples but rather created based on the patterns and structures it has learned during training,” ChatGPT goes on to explain in a “simple, but sophisticated” manner.

ChatGPT’s explanation of its inner workings not only defines what makes it run but also exemplifies just what generative AI can do: predict useable and coherent text (or pixels for images) based on large datasets. The large language models (LLMs) used to power generative AI to produce text could prove useful for the revenue cycle.

“By [leveraging an LLM], it lets you dramatically broaden the amount of training data that you have because you don't need to label anything anymore,” Varun Ganapathi, PhD, co-founder and chief technology officer at AKASA, told RevCycleIntelligence. “You have a ton of data from literally all of the English texts generated as part of the revenue cycle and elsewhere. There is a ton of revenue cycle applicability.”

However, the state of generative AI is still in its infancy, according to Austin Brandt, co-founder of Long Tail Health Solutions.

“With any new technology, there's a hype curve, and we are in the hype phase now,” Brandt stated, referring to the Gartner Hyper Cycle, which describes the lifecycle stages a technology goes through from initial development to eventual obsolescence. There are five main phases: technology trigger, peak of inflated expectations, trough of disillusionment, slope of enlightenment, and plateau of productivity.

Expectations are inflating when it comes to generative AI, with media outlets and industry analysts picking up on the new technology and its potential applications. This stage spurs greater interest in the technology despite limited practical use cases.

Brandt explained that healthcare will go through the trough of disillusionment when the initial enthusiasm dies down as issues arise with generative AI’s application to revenue cycle. However, at that stage, “we will see what this actually means for revenue cycle management,” Brandt said.

Revenue cycle leaders went through this cycle with RPA a few years ago. The technology promised to eliminate waste and reduce costs, but many organizations encountered challenges in achieving these goals on a larger scale using RPA.

“At that time, it didn't really do much for us, and now we're at this point where there's more sane adoption, more thoughtful adoption of the technology, and RPA is actually really good at certain things, but don’t touch it with other aspects of the revenue cycle because it’s not effective there,” Brandt said.

“We need to learn from past mistakes and try not to get caught up in the hype and force generative AI where it’s not intended,” Brandt continued. “When you’re a hammer, everything looks like a nail.”

Finding the use cases for generative AI

Generative AI is channeling a bit of fact and fiction as the healthcare industry—like many others—is hitting peak hype. However, some obvious use cases are emerging in revenue cycle management. For example, many healthcare organizations have used ChatGPT to generate appeal letters after a claim denial from a payer. Prior authorizations are another obvious area ChatGPT has been able to help with since the LLM can parse through relevant medical and administrative information available on the internet.

But those are low-hanging fruits, according to Sunil Konda, chief product officer at revenue cycle management solution provider SYNERGEN Health. Generative AI beyond ChatGPT can tackle larger issues, such as ineffective front-end processes.*

“It’s very critical to have accurate data about the patient upfront because if you don't have accurate information about the patient, most likely, you'll end up with a denial or another issue, and it takes a lot longer to address those issues. You would spend a lot more money and effort to fix those issues,” stated Konda. “On the front-end process, there are areas where a generative AI can help, like data validation and scrubbing.”

Claim denial rates are rising, and about 82 percent of those denials are potentially avoidable. Registration and eligibility issues, medical necessity errors, authorization snags, and other front-end issues account for most claim denials, costing providers money, time, and resources to rework and resubmit claims.

Generative AI may prevent avoidable errors in data entry by scouring thousands to millions of pages in payer contracts, policies and regulations, and other text-based documents to identify missing information or potential mistakes. It could also be used similarly to optimize coding, Konda hypothesized.

Another meaningful application in development is revenue cycle communications.

“There will be obvious implications or uses in payer and patient communications,” Brandt explained. “We've been migrating towards chat-based interactions for a while, and there have been even vendors who've been really successful so far in having revenue cycle-type conversations through a conversational user interface or chatbot.”

“Generative AI is a new technique to implement a chatbot, and it can maybe accelerate the development and even widen its capabilities, but it's not fundamentally introducing a new tool,” Brandt continued.

Similarly, a generative AI chat-based tool can be used for revenue cycle staff training. The tool can answer staff queries faster than it may take a professional to research and find the answer.

These are use cases revenue cycle management vendors and providers are just starting to look at. But Ganapathi sees a bright future for generative AI in the revenue cycle space. The use cases for revenue cycle will come from where healthcare organizations “need to turn some unstructured data into structured data.”

“That's where these LLMs and gen AI will manifest most powerfully. It may not affect how the EHR talks to the payer website or insurance company. It more manifests wherever you need to translate from things humans do, like English, to some structured document,” Ganapathi said.

What generative AI can’t do for revenue cycle

The healthcare industry must build the use cases to demonstrate where generative AI can transform revenue cycle and where it cannot. A grey area, though, is revenue cycle communication.

ChatGPT has proven effective at answering patient queries, particularly around routine screenings and care. However, communications around patient financial responsibility and financial assistance require a human touch. These conversations may hinge on a human’s understanding of tone, and it’s this understanding that optimizes the patient experience. AI chatbots can’t provide that level of understanding, Konda explained.

Even some claim denials may be too complex for generative AI to solve despite its potential to improve denials management. Some denials may return with a denial code, but that is not enough information to explain why a payer denied the code. It may be a credentialing issue, Konda said.

“You fix the credentialing issue with the payer, and all those claims can get paid. Generative AI is not needed or will not be very good in terms of strategy around addressing major denials or understanding why certain types of denials are happening. I don’t think generative AI is there to understand that level of detail at this point,” Konda stated.

Generative AI also likely isn’t the answer to process or workflow improvement. For example, revenue cycle leaders may reposition denials management staff to the front end of the revenue cycle to tackle issues leading to denials on the back end. That requires new systems, new processes, and new methods of validation.

“Some of those areas require process changes or bringing in a new technology to streamline the data collection process. In those areas, generative AI will probably not be very good,” Konda said.

Change management may help more than AI in some cases. Healthcare, after all, is riddled with inefficiencies stemming from the overly complex system. Healthcare also has significant data problems. In a survey released earlier this year, about a third of healthcare providers said less than 76 percent of their data is accurate. Furthermore, AI in healthcare has also posed serious algorithmic bias, data privacy, and security risks.

Generative AI relies heavily on the data it is trained on. If the training data contains biases or errors, the generated content may reflect these biases. Biased or inaccurate information can have serious consequences for patients and revenue.

Biases or other noises in data can also increase the chances of “hallucinations,” in which the generative AI tool provides text or images that are not based on real-world information or not a faithful representation of the input data. Hallucinations can occur when the AI model generates information that it has not been explicitly trained on or when it misinterprets or extrapolates from the training data in an incorrect or fictional manner.

“Because of the sensitivity part, you have to set policies on how you use it,” Ganapathi warned. “Also, coach your teams on what they can use it for and can't, and I think the answer right now is mostly general information. It also has a chance to hallucinate; these large language models can make up stuff. So, their main use is when you can verify its output very easily.”

When will generative AI streamline revenue cycle?

There is a long road to healthcare for generative AI, and there will be several bumps along the way before revenue cycle teams use the technology to improve operations.

“Healthcare is not the fastest-moving industry. So, it’s definitely not going to be tomorrow when everything is going to change,” said Ganapathi. “There will be some grassroots stuff happening, and we already see that with doctors randomly asking questions to ChatGPT. But in terms of it becoming the de facto way of doing things, that will take longer.”

Technology adoption generally takes some time because of different policies and regulations, especially around the security and privacy of data being fed into AI models. Healthcare is one of the most heavily regulated industries regarding data security because of concerns about the nature of patient data. More than 39 million individuals were impacted by a healthcare-specific data breach in the first half of 2023 alone.

“We can’t apply large language models straight up without taking those considerations into account,” Ganapathi stated.

Technology companies will need to solve the data privacy problem if they want to bring their technologies to bear. However, this is part of the usual technology development and marketing cycle for companies selling to healthcare organizations and won’t act as a barrier to adoption of purpose-built, HIPAA-compliant generative AI tools for revenue cycle.

“These requirements are addressable,” Ganapathi explained. And they can be addressed relatively soon.

Technology leaders estimated between two and five years before generative AI will take off in revenue cycle. On the lower end of that period, healthcare vendors and providers will solve the low-hanging fruit (e.g., prior authorizations and appeal letters) using generative AI, Konda predicted. About three years later, generative AI is likely going to target more meaningfully aspects of revenue cycle like denials management, Brandt and Ganapathi agreed.

“And revenue cycle will be dramatically changed by this technology,” Ganapathi emphasized.

Conclusion

A new type of technology is just around the corner for revenue cycle management. Doctors are already testing general generative AI models like ChatGPT to streamline some of the more cumbersome tasks of healthcare administration and larger healthcare organizations like UNC Health are looking into purpose-built generative AI models to target administrative use cases.

There is still some time before generative AI will be widely available and/or practically used to streamline revenue cycle. However, vendors and providers seem to be all in on seeing how generative AI can optimize an overly complex system.

“We all know that what we’re doing in revenue cycle needs to chance, needs to evolve,” Brandt said. “We need to leverage technology more and more as it becomes available and push the boundaries because we can't hire our way out of this. We can't staff our way out of this. We can't process and prove our way out of this into new levels of profitability or profitability at all. Technology has to be part of the answer to this.”

*This article was updated on 10/20/2023 to reflect Sunil Konda's new position.