2022 Risk Adjustment Factor Scores: Changing Tech, Data Strategies Postpandemic

Michael Stearns, MD, CPC, CRC, CFPC, specialized consulting director of medical informatics, Wolters Kluwer, Health Language, breaks down how payers must approach risk adjustment factor scores differently after the pandemic disrupted regular patient care.

What challenges do payers face regarding risk adjustment? How did the pandemic impact these challenges?

The pandemic has resulted in fewer patient visits and procedures, especially elective procedures. That has placed vulnerable populations at increased risk for worsening of their conditions. Patients not receiving routine care are also at risk for developing new diagnoses and complications.

The earlier conditions are detected and managed, the better the outcome. For example, estimates indicate about 40% of patients with stage 3 chronic kidney disease are not even aware they have this condition. If they go a year or two without being seen by health care providers for this condition, the patient's kidney function is at risk for further deterioration.

Chronic conditions tend to evolve over time, so risk adjustment factor (RAF) scores from 2020 and 2021 may no longer be accurate. Patients may not have seen their specialists, which could result in a different RAF score that is not truly reflective of the actual cost of caring for members in the payment year.

It will be a challenging year for risk adjustment stakeholders.

What technology might payers use to approach risk adjustment more efficiently?

Basic analytics have been used for years, and they can play a major role in identifying patients who have underlying conditions that have not yet been reported in the targeted year.

For example, patients that have been treated for breast cancer may be placed on a medication such as tamoxifen to reduce their likelihood of recurrence. They are still being treated for breast cancer, however, which is a reportable condition for risk adjustment.

If it is not documented, which happens not too uncommonly, that medication can serve as a "clinical indicator,” or a clue there is an underlying diagnosis that has not been reported. There are many thousands of examples of clinical indicators that can be used to smartly recognize underlying conditions that have not been reported.

Another topic that is at the forefront for risk adjustment is natural language processing (NLP). Traditionally, the level of NLP's accuracy in clinical medicine has not supported real-time care decisions at the provider level, such as when there is evidence a specific medication should be prescribed. However, NLP can improve the accuracy and efficiency of coding medical records.

The computer can churn through thousands of documents and look for specific data it has been taught to recognize. NLP applications have been variable in their ability to identify conditions, but they do much better if they use a rich source in medical terminology and are customized for the risk adjustment use case.

Machine learning is another technology that should be considered. Machine learning uses neural networks to identify and correct errors, and eventually the machine gets smarter and smarter. It is a form of artificial intelligence, so it is quite popular right now.

The challenge, however, is you can have the best tools—the best NLP, medical terminology, and resources—but the pieces may not fit together well. You must spend a great deal of time fine tuning to get the desired level of accuracy, and that will continue to improve over years.

This does seem to be the current direction: increasingly using machines to help with the efficiency and accuracy of coding.

What kind of data should payers be looking for to accurately adjust risk profiles?

There are a growing number of potential data sources, which is potentially challenging.

Claims data is straightforward and has been used for many years. If a patient had a condition reported in a prior year, but has not reported this year, that is a red flag that the condition may still be present.

The example commonly used is amputations. If you have a lower extremity amputation, such as the removal of a big toe, it is possible that was not included in the claims data the following year, so the diagnosis may go unreported. The IC10 code for amputation status could have been reported in the past, and then claims data can be used to identify related procedures, such as the removal of the toe and medical devices like prosthetic limbs. All those data come together to help you identify suspect conditions, allowing you to improve your performance and RAF score.

Another well-organized source of information that may not be used as extensively is lab data. Certain lab values may be strongly associated with a specific condition and may indicate a test needs to be repeated or a diagnosis can be confirmed. This is huge because it means the patient can be appropriately cared for, and any underlying conditions can be addressed to improve the patient's long-term outcome.

NLP can identify supporting documentation in narrative records. Supporting documentation is required, of course, when you submit diagnosis codes for Medicare Advantage, and all IC10 coding. This is commonly referred to as the retrospective model and is also applicable to the concurrent model for reporting.

The retrospective model involves identifying diagnoses supported by documented evidence the condition was evaluated and managed during an eligible encounter. There is a lot of criteria, but machines can select eligible documents and specific data elements within those documents, which will help you identify an unreported condition.

The prospective model, to review quickly, is where an unreported diagnosis in the targeted year may be present but not adequately documented by the clinician. Say the diagnosis is only found in the problem list and by the organization’s policy is not reportable, or the diagnosis is not even mentioned in the record. But there is evidence the diagnosis exists, such as those clues I mentioned we can pull from any source of documentation.

For example, there is a diagnostic MRI of the chest. You cannot report diagnoses from that type of report for Medicare Advantage, but it may have a finding that is strongly suggestive of chronic obstructive pulmonary disease. Some findings, such as hyperinflation, may be diagnostic by themselves but require a future doctor’s visit to be reported. We would consider this a high-confidence clinical indicator, and that clinical indicator would then be presented to a coding professional, who will determine if the evidence is strong enough to share with a provider during a subsequent care visit. A provider can then look at that information and potentially confirm a diagnosis. It also gets the patient engaged in needed care.

How do risk adjustment strategies impact costs?

Accurate RAF scores are key to supporting the risk adjustment model. The cost of implementing these strategies, including technology, analytics, and other programs, needs to be taken into consideration. Each strategy must have a proven ROI, which depends on the setting of care and other specific variables.

The efficiency of coding is also being evaluated. Reviewing every chart is impossible without technology. Technology can do a first pass or second pass review and identify codes, which need to be validated by a human reviewer. But the fact that they can take 1000 charts and point to 5 or 6 that have relevant information, as opposed to the reviewer going through the whole sample, has obvious value.

As the health care landscape continues to change, what should payers should keep in mind regarding risk adjustment?

At RISE West 2021, the people I spoke with said there is growing concern regarding the Office of Inspector General investigations and Risk Adjustment Data Validation audits. The concern was over what is considered supporting documentation. It can be a subjective decision that is made by coding professionals, in good faith. But we lack strong guidance, in my opinion, as to what exactly represents supporting documentation.

My understanding is that acronyms like MEAT, which stands for monitor, evaluate, assess, treat, are created by the industry. They are very useful when determining what is and is not supporting documentation, but it is not 100% reliable if you were to undergo an investigation.

Given the focus on compliance, organizations may need to revisit their local policies to see if they have any concerns or gray areas. Hopefully we will get further guidance from the powers that be. We do have guidance for the lifelong permanent conditions list, and that was a breath of fresh air to me, because it listed specific conditions that do not require supporting documentation.

To my knowledge, that level of clarity does not exist in Medicare Advantage, which creates some confusion for providers. I have heard some payers say trying to teach providers how to code really depends on what is required, whether they are a Medicare Advantage or commercial payer.

Is there anything else you would like to add?

One thing I think is neat and draws from my background, too, is the use of interoperable data in risk adjustment, which is critical for patient care. The government has been focusing on improving the ability for systems to communicate with each other, and payers are being asked to participate as well.

There will be a lot more information available through health information exchange platforms that can be used for risk adjustment, from lab data to claims data to documentation that has evidence of a condition. This information is a rich source.

There is also a renewed focus on chronic conditions. There is a Chronic Care Management program, which Medicare Advantage and some other organizations have embraced. It rewards things like prompt follow-up with a patient after leaving the hospital.

Technology can also be used to ensure the patient obtains their prescriptions, which is a common issue with patients because they cannot afford them or do not have transportation, etc.

It is also worth mentioning social determinants of health (SDoH). These are factors that may influence the quality and cost of care being provided to patients. These include food insecurity, economic insecurity, transportation issues, and a wide range of other things. The list is rapidly growing.

I know Medicare has been looking at SDoH closely for RAF scores. SDoH are also being used in other programs, like the Merit-Based Incentive Payment System (MIPS), which Medicare providers are required to enroll in. It would be good for payers to become comfortable with that approach.

About Dr Stearns

Michael Stearns, MD, CPC, CRC, CFPC, is a physician informaticist with over 20 years of experience in health information technology. His work has involved electronic health records, telehealth, virtual services, quality reporting, risk adjustment, health information exchange, clinical terminology development, standards, and billing and coding compliance.

Dr Stearns is currently the specialized consulting director of medical informatics at Wolters Kluwer, Health Language. Previously, he served as the international director of SNOMED International; founding board president of the Texas e-Health Alliance; cofounder and lecturer at the Health Information Technology Certificate Program, University of Texas at Austin; cochair of the 2020 American Health Information Management Associations’ Technology and Innovations Practice Council; and other roles.