NLP is an AI technology that is being used in healthcare IT for clinical documentation and medical coding. For medical coding, the program identifies diagnoses codes for HCC risk adjustable categories and flags it for a medical coder to review.

In robust medical charts that span up to thousands of pages in length, this enables coders with an automated way to identify diagnoses codes for review, hence increases speed, efficiency, and output. Academic research has found NLP increases medical coding productivity by 15-20%.

After the medical chart is reviewed by a medical coder, the chart goes through a pre-submission QA process for accuracy and compliance review. In some cases the chart will go through the NLP program for a second pass to identify additional insights and potential missed opportunities.

Tune in to this episode to learn:  

Host: Today, we’re talking about NLP for Coding and Compliance with Kristi Reyes. Kristi is the Director of Risk Adjustment Coding Operations. She leads a team of 250 medical coders from various aspects of risk adjustment—including retrospective, prospective, CDI, QA, risk mitigation, and RADV. Kristi holds certifications in Certified Professional and Outpatient Coding, Risk Adjustment Coding, and Auditing. Welcome Kristi.

Kristi: I’m excited to be here!  

Host: We’ve heard a lot about AI lately it’s Chat GPT. We’ve seen everything from TikTok to TED talks on the use of AI models. So, let’s talk about the use of an AI model called NLP, Natural Language Processing. Can you tell us about NLP and how it’s being used in the medical coding world?

Kristi: Yeah, so NLP is an AI technology program that is being used in healthcare IT for clinical documentation and medical coding. For medical coding, the program reads human language and text in medical records based on algorithms that match diagnoses that fall into HCC risk-adjustable categories.  

As an example, the NLP program will scan through the medical record to identify a specified set of words. Let’s say diabetes. The program will identify any instance of diabetes, whether it’s specifically documented as diabetes or documented as a myriad of potential acronyms, jargon, or abbreviated formats like DM or DM Type 2.

This has a number of benefits. So, the first benefit is speed and efficiency. In robust medical charts that span up to a thousand pages in length, clearly, a computer can scan that faster than a manual process. Additionally, the sheer volume of data requires automation to parse through in a timely manner.  

The second benefit is quality control and reduction in human error. Key indicators and diagnosis codes can easily be missed. NLP provides quality assurance so we can maintain a 95% accuracy and completeness standard. Accuracy means the codes abstracted are appropriately documented and supported in the medical record. And completeness means that all the diagnoses codes are appropriately documented and supported in the record were captured. In some instances, there can be inconsistencies between coders and how they interpret medical language, so NLP helps to provide an objective base. Coding can be a gray area. If two different coders look at the same chart, they may each conclude two different coding scenarios. One coder might find one thing, another might find something else, and a third coder might miss something.  

NLP provides quality control and assurance that the code will be identified for review by a certified coder. Now, whether it’s risk adjustable or not… if it’s not risk adjustable, maybe we can use it to provide a clinical opportunity within the context of a prospective program.

Going back to the benefits. It’s really an indirect benefit, and that’s continuing education.

As NLP programs identify and direct human coders to keywords and diagnoses codes, coders have the opportunity to observe and reinforce their own knowledge and skill set. [For example, the NLP is showing me, Purpura?, I might say, “I didn’t know that risk-adjusted. Or I forgot that risk-adjusted.” It’s this kind of support that increases production.

Host: Diving in deeper, what’s the relationship between NLP and coders? What does the process look like?  

Kristi: Good question. So, let’s say the coder logs into the medical record, and if they don’t have NLP, they’re starting from scratch—having to review the whole medical record. Scratching their eyes to find these keywords. But, if they have NLP, the NLP is going to bring them straight to that page number and, in some cases, even highlight the keyword they’re looking for. For example, in risk adjustment, we have a responsibility to identify and capture diagnosis codes to correctly reflect the burden of illness for members. So, the NLP for our clients is preset and designed to thumb through and find specific chronic conditions (HCCs) that must be documented and reported annually for risk adjustment programs. We’re seeking codes that accurately and completely capture the chronicity of illness for our members to make sure we document the burden of illness for that member population. So, at this point, we’re talking about First Pass NLP. As a first step, the medical coder can get through the medical record a lot quicker. The program might say there’s an instance of diabetes on page 1, page 10, page 50, and page 1233. Now, the coder can go to those pages and apply their coding and compliance guidelines to see if they are able to extract any of the documented conditions. The great thing about NLP is it can be used for all kinds of coding programs: it can be used for HEDIS, CDI, provider education, SDoH, and even suspecting logic and analytics.  

Host: So, it sounds like NLP acts like digital sticky tabs in the medical record?  

Kristi: (Laughs) Yes, exactly. And I think that’s an important distinction to make. NLP is not replacing the coder; it's flagging what needs to be looked at by the coder. It’s very much a tool. It helps us, coders, hone in on keywords. You know, some charts are only 10 pages, while other charts can be thousands and thousands of pages. NLP supports the coder and ensures the maintenance of a 95% accuracy and completeness standard in a very compliant process.

Retrospective risk adjustment is looking to close historical gaps so we have an idea of what conditions the patient has, and NLP models are being implemented in this environment also to flag suspected gaps. But we also need to account for newly developed conditions, so we don’t necessarily know what conditions we’re looking for. The basis of retrospective coding programs is to not only capture historical conditions previously reported, but also to capture diagnoses that are newly developed and documented for the patient. This is likely to take the form of incremental pieces of evidence. NLP is a really good tool for highlighting what we need to look at, allowing the coder to quickly find the information to make a decision about whether the condition is documented and supported in the medical record.

Host: Let’s zoom in. What’s the process look like? What’s the lifecycle for the medical chart from first pass to submission?  

Kristi: Like we just touched on, the chart goes through NLP First Pass. This is done to pre-identify which charts contain certain keywords for diagnosis codes. It segments the charts by code type so we are reviewing and capturing according to the specific coding program parameters, and it identifies charts that may need a higher level of scrutiny to reduce regulatory risk. After First Pass, the chart will go to a medical coder. The medical coder is going to review the entire note, all pages, including everything that First Pass flagged—they’re always looking for the best note, the note that has the best-supporting documentation, or what the industry calls meat. If the NLP has directed the coder to ten different instances of diabetes, that coder will review the ten instances and select the record that has the most supporting documentation. Maybe the best evidence is where an A1C is documented; maybe they’re refilling insulin or Metformin. Maybe there’s a referral for podiatry or ophthalmology for a diabetic check.  Or maybe there’s an order for a lab for A1C or glucose. Coders are going to be looking for the record that has the most meat, so there’s confidence when it’s sent to CMS for risk adjustment submission, lowering audit risk.

Host: So, NLP finds all the evidence available, and the coder selects the most convincing evidence…

Kristi: Exactly. NLP allows us to get there quicker. If you don’t have NLP, you start on page 1 and have to get to a thousand pages relying on coder eyes alone, and as we know, each coder may see the record differently, and things can get missed. But if NLP is driving the coder to those specific insights, you’re going to save a lot of time while improving your completeness—two sets of eyes are better than one!

Host: What’s the next stop for the chart after the medical coder?  

Kristi: Then the medical chart goes through our pre-submission QA process for an accuracy and compliance review. A QA Coder reviews each abstracted diagnosis code and ensures it meets client, coding, CMS, ICD-10, and compliance guidelines for submission. Then, the final step is for the chart to go through a final manual review with a medical coder before it’s submitted to CMS or the client. For our clients, we also offer a second-pass NLP solution. These are records that have already been reviewed by a medical coder and are then run through a separate NLP technology to identify and additional missed insight that would require additional coder review. This multiple approach of NLP utilization can be very helpful, as we discussed different coders seeing things differently when reviewing the same record.

Host: At this time, does NLP have any limitations?  

Kristi: Yeah. So, we have 2 ways we access medical records. One is scanned PDF, where a retrieval specialist goes out to get those records. And then there are EMR records. NLP cannot get run through an EMR that a coder is accessing remotely. So, we obtain these scanned records to run through the NLP technology. There are currently initiatives that are in the works to be able to use CCDA (Consolidated Clinical Document Architecture) to format the EMR record into a PDF so the NLP program can run. If, for some reason, there’s a formatting error and the PDF doesn’t contain the minimum DPI, dots per inch, then that’ll obstruct the NLP program’s use of optical character recognition (OCR) to scan the medical record for keywords. Yet another reason why we need a human coming behind the program for quality assurance.  

Another example of why coders are needed to support NLP is that the coder has to look at all the flagged instances; let’s just use diabetes as an example. They need to determine whether there is conflicting information in the medical record. Let’s say on page 4 of the visit; there’s an instance of diabetes; the provider has prescribed Metformin, running an A1C. But then we scroll back up to an HPI on page 2 and we see the provider notes indicate the patient is there for pre-diabetes. That’s conflicting information, and we can’t abstract that for retrospective. We can use that for prospective reviews and send a query to the provider for clarification so it’s still useful.

The documentation and capture of active cancer also require a coder. Physician documentation is a challenge here—determining what constitutes active cancer versus a history of, and what is risk adjustable. Having NLP identify the diagnosis codes is helpful, but we have to have the coder review to make sure that there is supporting documentation that there is active cancer and active treatment. Another diagnosis that’s been a hot topic for compliance is acute stroke in the outpatient setting. We just don’t see that. So, if NLP identifies a stroke and the place of service 11 and it’s a provider note, we know that it’s a history of stroke, and we don’t abstract that.  

Another diagnosis that may slip past the NLP program is PVD, which can stand for 2 different things—Peripheral Vascular Disease, a cardiovascular condition, or Posterior Vitreous Detachment, an ophthalmic condition. PVD is also synonymous with vascular claudication. And there’s also neurogenic claudication, which is a completely different condition. This is especially important in risk adjustment coding programs, where certain diagnoses are used to calculate the member’s burden of illness. So, when the NLP identifies claudication, we have to make sure to capture these conditions with accuracy.  

There are so many examples of why coders are essential for compliance.  

You know, the NLP program has its limitations; it needs quality assurance checks, too. So, there has to be a way to provide feedback to the NLP program. We do this by tracking diagnoses identified by NLP that are abstractable versus those that are not. It is important to also have a feedback loop with NLP, especially if you are using a vendor. Regular detailed adits of NLP findings and omissions all allow for continuous improvement. Another example of a limitation where feedback is essential is incorrect text mapping to something that’s not relevant. To use an earlier example, neurogenic claudication should not prompt a result for vascular claudication. Or we’ve seen instances where the patient has diabetes, but it’s actually a family history. Then, we also have to ensure the NLP doesn’t identify something that’s already been resolved. We want to be sure insights that aren’t capturable are excluded. NLP is great; it’s still saving mountains of time, making the process more efficient, and ensuring higher levels of accuracy and completeness, but there still has to be continuous improvement in the NLP rules engine.

Host: I suppose the system has to get updated fairly often as new codes emerge and regulatory updates come out. How are updates accounted for?  

Kristi: Every year, coding updates have to account for ICD-10 changes, the V28 model comes into play for Medicare Advantage risk adjustment, new SDoH guidance emerges, regulatory changes from CMS, and industry trends. Each plan at that point has to decide on the parameters and strategic direction they want to take within their risk adjustment program and we set the NLP rules accordingly. So, each plan may have slightly different NLP rules based on whether that plan wants to take a more conservative approach or take on a bolder approach; they can make amendments to the coding guidelines accordingly.  

Our NLP casts a wider net. We want to oversaturate and not undercut because the goal is to identify any instance we can possibly find that will accurately support and document the member’s burden of illness. We want to pull in every instance and then have a coder go through it and make sure it’s compliant with the coding guidelines and select the best evidence.

Host: What’s the average amount of time saved from using NLP?  

Kristi: Great question. First, that’s going to depend on the chart length. If it’s a 2-page chart versus a 1000-page chart. There’s academic research that shows it can contribute to about a 15-20% increase in productivity. And as I’ve mentioned before, we have a 95% accuracy and completeness standard. So, between the NLP and our coders, we’ve got a good process that outperforms the industry.  

Host: Kristi, thanks so much for joining the show today. We learned a ton about how NLP enhances the medical coding process for risk adjustment. To our listeners, thanks for tuning in. We’d love to hear from you. Leave a review on Apple Podcast or Spotify, and follow the show to get notifications when new episodes are released.