Summer 2024 Issue

Autonomous Coding and Data Integrity
By Susan Chapman, MA, MFA, PGYT
For The Record
Vol. 36 No. 3 P. 10

Accurate coding is fundamental to data integrity. Computer-assisted and autonomous coding are helping.

A crucial aspect of health care is data integrity—ensuring that health information is high quality, accurate, and complete. Not only does this information affect the immediate care patients receive, but it also can influence their future health care. Additionally, high-quality data can be integral in implementing positive public-health efforts and reducing health care costs.1

Computer-Assisted Coding vs Autonomous Coding
A significant aspect of data integrity is accurate coding. Over the past decade, coders have been able to rely on computer-assisted coding, a process by which they can select from a range of codes suggested by software based on health record information. Although computer-assisted coding has played a critical role in streamlining the coding process, it still requires not only a coder’s attention but also their time and has not necessarily improved overall accuracy. Now, autonomous coding, with which a patient encounter can one day be entirely coded by computer, is showing the promise of greatly enhancing data integrity.

“Computer-assisted coding doesn’t necessarily improve data integrity in isolation,” according to Andrew Lockhart, cofounder and CEO of Fathom. “Its quality hinges heavily on how accurate the coder is who’s using the tool because what it’s doing is pulling up an encounter and suggesting many codes for the coders to sift through. They then select or delete the suggested codes, which is essentially a manual process,” he explains.

Even though this form of coding still requires human intervention, computer-assisted coding can streamline the coding process by augmenting coder expertise with rules-based software, which supports code assignments by analyzing elements like clinical documentation, diagnosis procedures, and treatments. “What typically happens in the absence of having software present that information to a coder is that they have to research the information and add any potential missing codes before they finalize the record for adjudication,” says AGS Health’s general manager of coding automation, Conrad Coopersmith. “Skill sets differ from coder to coder with respect to their level of knowledge. The upside of having technology in this context is the impact on efficiency, because the computer is assisting with the analysis itself. It’s generating code suggestions and enabling a much quicker resolution, which then creates more leverage for coders to focus on cases that have greater complexity.”

Computer-assisted coding has been available in health care for about 10 years. Over that span, the technology has improved workflow and paved the way for autonomous coding, a form of artificial intelligence (AI) that automatically assigns codes without coders’ needing to intervene and make coding decisions. Not only is a computer automatically assigning codes based on information from the EHR, but it’s also learning and applying that knowledge to future scenarios. “The model is analyzing voluminous amounts of data to be able to learn from that data over time, and that includes everything from patterns to medical practices to coding guidelines. There is a difference between harmonizing data to make people more efficient and having a system that is dynamically learning over time to automate the direct-to-bill process,” Coopersmith says.

Autonomous Coding and Improving Data Integrity
Fully automating the coding process can have inherent risks in that organizations may not have enough information to use a machine-learning model with confidence. They can also be affected by the quality of the data that goes into the system from the start. However, the process of manual coding is time-consuming and often error-prone because people can enter incorrect codes, which then leads to claims denials and provider organizations suffering financial losses.

From a data-integrity perspective, organizations can realize gains with the systematization of processes. Automation can allow for vast amounts of structured and unstructured data to be processed instead of that work being performed by a person. In turn, that automation can reduce the likelihood of error. Automating the coding process can also create consistency with respect to how codes are assigned, and when the machine is trained over time, accuracy can improve.

To address accuracy concerns, Diana Ortiz, RN, JD, CCDS, CCDS-O, senior manager of global content at Solventum (formerly 3M Health Care), states, “We advise clients to set confidence thresholds in order to test the system and maintain some oversight on the journey toward automation. We don’t recommend just ‘turning it on’ and having a black box send information out. Doing that creates a risk of data-integrity issues that can come up. Instead, we recommend starting the process with service lines where we know that the documentation is good and there are high metrics around tools like computer-assisted coding. That is low-hanging fruit where there is less risk but more confidence. Then, we’re reaching a 95% threshold of accuracy, which is good enough for an organization. From that point, they can consider examining the other 5% or creating an auditing process for spot-checking until they really feel comfortable. There are a lot of checks and balances that can be put into place. It’s definitely a topic that is top-of-mind within the industry, too. And, while there is a push for automation, there is also the goal of minimizing the risk around it.”

Lockhart agrees with Ortiz that there must be a system of checks and balances to verify data accuracy. “We encourage clients to evaluate performance, which includes our doing audits and providing reports, but we also say to treat AI like you would a coder on your staff,” he explains. “Have checks in place and regular audits to make sure that the system is performing the way you’d like it to perform. One of the unique benefits of autonomous coding is that you’re able to validate the system at enormous scale prior to going live. One way to do that is by coding side-by-side, a month of coding with AI alongside manual coding. The clients can then see for themselves the difference. When coding leadership can do an audit of these two processes, they are always surprised to realize that they prefer autonomous coding over their existing process. A process like this gives decision-makers a chance to vet the system up front and demonstrate its performance.”

Over time, the need for spot checks, auditing, and other processes to evaluate an autonomous-coding system may become less frequent. “If you look at advanced technologies in this arena, multiple automated controls exist, and they exist for oversight as checks and balances,” Coopersmith says. “What’s important about that is they have to exist in the same system that the coding is being performed in to ensure that quality is being met. Newer technologies have the capability to audit work prior to claim submission, which ensures that the appropriate codes are being captured. Another capability is auditing retrospectively following claim submission to confirm reconciliation as well as expected reimbursement. Moreover, enhanced technology has the capability to audit the auditor. All of those things are requirements, and they’re going to continue to exist.”

The Need for Queries
Just as some form of verification will continue to be part of the autonomous-coding process, queries will still need to take place, but the need for them could be reduced. “With AI, clinical documentation and codes are automatically and continually being analyzed, and they’re being flagged against past performance,” Coopersmith explains. “The system enables swifter resolution and improved accuracy over time as new information is being introduced and things are happening in real-time vs retrospectively. I think the impact over time as the systems get smarter will lessen the need to query but there will always be the need for human intervention.”

Lockhart agrees. “There might be something going on in the actual record at the time of treatment that would necessitate a query. For example, an emergency department visit where things are occurring across multiple physicians who are performing multiple functions. All of that needs to be bundled into a singular encounter. It’s normal for an autonomous coding system to receive all the encounter data and then flag a missing signature or history. That would then require the clinician to complete the documentation. I would say, though, that it is very exciting, as more of the health care system moves towards AI transcription, we’re going to see less of those cycle-backs. As AI becomes more present in both coding and documentation, the amount of resources dedicated to finding and resolving those discrepancies is going to go down in a way that I think is much bigger than most people anticipate.”

Financial Benefits
Reducing costs, accounts-receivable (A/R) days, and denials is among the largest concerns of early adopters of autonomous-coding technology. “Denials is a lagging indicator, so while we see the potential for them to be reduced, the information isn’t there yet,” Ortiz says. “Organizations are positioning some of their coding or documentation staff to help work on denials as soon as they come in, whether it be with appeals or stop gaps to see if there is something they should look for beforehand. They’re implementing a lot with existing staff, and the challenge is how best to utilize that staff, which is where automation can eventually pick up in some areas. There is this expectation in the industry that we’re not going to hire or train more people, and automation will solve some of the challenges that go along with that.”

“By virtue of automating code assignments, coders are able to transition their focus to exception vs inspection,” Coopersmith adds. “With an exception-based approach, analyzing only those things that come up, providers are able to drive denials down and reimbursement up because they are getting it right the first time.”

Lockhart believes his clients are experiencing dramatic financial benefits, including reductions in cost, A/R days, and denials, as well as significant revenue uplift. “With a manual workforce, you can be far more subject to seasonality, people getting sick, losing a staff member. All of those things can drive up your A/R gauge overnight, whereas an AI system is going to consistently be performing,” he says. “On denials, there’s a lot of low-hanging fruit we find with clients just in terms of cleanly implementing coding guidelines and then systematically enforcing them. The second trend is overall denials are going up, and so the net impact is largely in the payer’s hands. But with automation, coding-related denials should certainly plummet. Lastly, there’s revenue uplift. By and large, humans tend to under code, particularly E/M [evaluation and management] levels, so switching to autonomous coding can not only increase coding accuracy and compliance but also result in pretty significant revenue uplift.”

Accessibility
AI-driven data-integrity solutions can be costly to purchase and implement, which may mean such systems are not necessarily accessible to smaller or rural health care organizations. “Currently, automated systems do align better with larger organizations,” Coopersmith says. “My belief is that companies with technology and manual coding services are going to be in a unique position to bend the cost and adoption curve due to the institutional data and learnings they already possess from performing the work. In those instances, there will be multiple options that result in savings and efficiencies passed on to the provider regardless of the provider’s size.”

Lockhart adds that the immediate return on investment (ROI) of an AI-driven solution could mitigate the impact of purchase and implementation costs. “For our clients, we try to be ROI-positive from day one. Implementation fees are pretty nominal if at all. Then, on a per-encounter basis, they’re going to realize better accuracy, which means they should see a lift that can vary depending on what the client had in place before. In time, smaller provider offices and rural hospitals will be able to access these solutions generally through an integration into their EHR,” he says.

AI’s Capabilities and Capacity for Growth
Because the success of any AI-driven solution depends on the integrity of an organization’s documentation, aspects like a system’s efficiency, speed, and scale are all dependent on having strong underlying data. “The data piece has to be solid. If you’re trying to employ an autonomous system without a really good documentation plan and framework, it’s only going to be as good as what you put into it,” Ortiz says. “It does also depend on scale. If you’re only feeding the system a limited set of documents, it really can’t fine tune and learn from itself in a way that proves better accuracy over the long term. I would not say that is a pitfall, but to do this well and right also requires scale, which is an important thing to consider when selecting a vendor.”

Autonomous coding is affecting health care in broader ways, with gaps in AI’s computing capabilities decreasing between large and small organizations. In coding, AI-driven solutions are transforming the coder’s role from that of manual coding to one of more analysis and audit. “I think we’re at this crucial time in the industry where AI is not just in this one space; it’s across health care—driving new medicines and new devices at an exponential rate,” Ortiz offers. “So, I think we all care that that the data are reflective of the things that will go into AI’s models that will subsequently drive many things across the industry.”

— Susan Chapman, MA, MFA, PGYT, is a Los Angeles–based freelance writer and editor.

 

Reference
1. Data quality and integrity: advocacy. AHIMA website. https://www.ahima.org/advocacy/policy-statements/data-quality-and-integrity/