Risk Adjustment Coding Gets a Technological Assist

How a California physician practice designed a system to catch coder errors without burdening doctors.

Technology continues its climb toward what we would consider intelligence. Your average electronic health record is fairly stupid, being able to sort things in rows and columns, and find character strings. Analytics software adds more sophisticated algorithms that can stratify patients by comparing their symptoms to known risks for developing more serious adverse events.

Jennifer Pereur

In the nearly three years since I wrote about IBM's Watson and its attempt to move health IT into an era of what has become known as cognitive computing, I have seen scattered attempts to further explore its possibilities. Some concepts, such as rules-based engines, expert systems, heuristic algorithms, and other machine learning, have been around in one form or another since the 1970s. More recently, technologists have incorporated these components into the broader knowledge representation systems that cognitive computing encompasses.

I recently spoke with an independent practice association in northern California that seems to be pushing cognitive computing into an area I had not heard about before, this time not strictly to explore clinical research goals, but to achieve real benefits to the group's bottom line.

Jennifer Pereur is director of government programs at San Ramon, California–based Hill Physicians Medical Group, a provider network of more than 3,800 primary care physicians, specialists, sub-specialists, and consultants that manages 300,000 patients lives at risk under contract to Medicare Advantage, managed Medi-Cal, and commercial payers, as well as "some PPO work as well," Pereur says.

The Medicare Advantage dollars that CMS pays the group are risk adjusted, Pereur says. "If somebody has a lot of conditions, they get a higher risk score and, therefore, there is more money to take care of them and provide care for them," she says. "If it's a healthier member, then there are fewer premium dollars."

Therefore, Hill Physicians employs many different mechanisms to make sure it is getting accurate data on the health status of its patients. "Not only does it help us with accurate premiums, but it also helps us to identify members who might be good candidates for different disease management programs," Pereur says.

The difficulty is getting physicians to document data in ways that are easily sharable with other physicians—a challenge facing all of healthcare today. Too often, the data that ends up getting shared is that data that ends up being coded on a billing claim. "We lose data along the way," Pereur says.

For example, one physician might diagnose that a patient is depressed, but that chart note does not find its way into structured data entry. This works against Hill's mission of having an accurate risk adjustment to present to payers, and it also does not help connect dots between medications that patient may have been prescribed for the depression, and how those medications might affect overall patient care.

To address this, Hill Physicians turned to technology from Apixio, whose newly announced cognitive computing platform, Iris, combs through patient chart notes to identify diagnoses. Coders log in to the Iris-powered HCC Profiler application and are presented with a series of its data discoveries, which the coder can then choose to accept or reject. Accepted suggestions create more accurate structured electronic health record data. Suggestions coders end up rejecting give Apixio suggestions on how to fine-tune its cognitive computing algorithms. Out of 100 different documents, Apixio may only present two documents with potential risk-adjusting conditions to be reviewed, Pereur says.

Having seen a demo of how this works, it's possible for me to see how combining the kind of natural language processing that we've seen in many other healthcare IT scenarios with the workflow requirements of coders is a potent union, particularly for an organization such as Hill, where so much is riding on getting the right care to the right patient.

"Now everyone who goes in [can look] at that holistic picture of the member; multiple people can now see that information," Pereur says. "We've taken it out of the chart, and we've put it in a place where we can do our care management."

Because the main action takes place behind the scenes with Hierarchical Condition Category coders entering the Risk Adjustment Factors that CMS and other payers require—not putting the suggestions in front of physicians at the point of care—Apixio's process has had minimal impact on actual clinician workflow since being deployed at Hill Physicians nearly two years ago, Pereur says. Prior to this, the burden of directing people to comb through the charts manually made it cost-prohibitive, she adds.

I asked Pereur if Hill Physicians also has a data warehouse that could achieve the same result. "What's in our data warehouse predominantly is administrative data—payment claim data, lab data, pharmacy data," she says. "We don't have all of that rich data that lives within the electronic health record." Apixio can grab data from any data source, not just the EHR, she adds.

Apixio can also spot codes entered incorrectly into the medical record, Pereur says. "CMS has stiff financial penalties for submitting data that's not backed up in the chart," she says. "So by going in and combing the chart, it also allows us to not only find data that we didn't have, but confirm data we did have, and potentially remove data that's not substantiated."

Hill Physicians says it is seeing an ROI of 7 to 10 times on its expenditure on Apixio technology.

While the unstructured portion of the medical record remains problematic, and smarter processing can't cut down on such information's verbosity, duplication, and possible contradictory elements, we might be seeing a move here toward a kind of machine learning that puts us on the path to clinical documentation done right, or at least better than it has been. And I like the idea of doing so without retraining physicians.