Radiology Coding Alert

Choosing the right artificial intelligence (AI) coding system for your healthcare organization isn’t like walking into a big box store and grabbing the product off the shelf. Medical coding contains specific guidelines, documentation nuances between providers, and regular training of the technology to ensure the outputs are accurate.

In this two-part series, you’ll learn the key areas that will help you identify the ideal AI coding solution for your healthcare organization.

How Accurate is the Technology?

The AI coding system’s precision is by far, one of, if not the most, important point to analyze when choosing a solution.

When a company develops an AI coding solution they have a level of confidence in its accuracy. For example, if the vendor claims the technology successfully handles 85 percent of your claims, they are 95 to 97 percent confident that the system processes that 85 percent.

“But if there’s an issue, you’re going to have to address that, and change all those claims from the back end. There’s still some oversight that’s going to be required when assessing a solution’s track record,” said Jessica Miller, MHA, CPC, CPC-I, subject matter expert at MediCodio, during her session “Successful AI Integrations: Case Study” at AAPC’s HEALTHCON 2024.

As you’re evaluating vendors to make sure their product is on the same page as your goals, remember that the technology’s accuracy will grow and learn with your inputs as well. A software that can accurately handle 85 percent of your claims upon deployment could reach 95 to 97 percent accuracy over time based on your feedback and additional data learning.

How Fast Does the AI Work?

Unlike your car or an internet speed test website, it’s difficult to estimate how quickly AI evaluates the medical record, deduces which codes to use, and generates an output for your review. “Some of these processes can actually happen in less than 20 seconds, but it’s just a matter of how much data we’re pulling. If we’re going to pull large record for inpatient, that’s going to affect the bot processing time. It’s going to have more records,” Miller said.

As you’re evaluating AI solutions, you can consider asking the vendor typically how fast the technology should process a record and what is a realistic processing time for the amount of data you want the software to review.

Does the Software Integrate Into Your System?

Every practice and organization have an electronic health record (EHR) system, and no one wants to change that system just to add AI into the mix. Completely changing your EHR system will have a high financial cost, your team will need to be trained on the new system, and productivity will slow while the practice gets acclimated to it.

Instead, you’ll want to look for an AI coding solution that integrates seamlessly with your existing system. Simultaneously, as you deploy the technology with the EHR, you’ll need to perform the correct data mapping. The AI software may be trained for certain data points in a medical report, such as “procedure description.” However, if your EHR uses “procedure” or “operative procedure,” then you’ll have to customize the software to search for that specific information to ensure the correct analysis.

“You’re always going to have to be checking that integration and those field mappings are in compliance or in congruence,” Miller added.

Does the Algorithm Accurately Interpret Documentation?

For the AI coding solution to meet your needs and help streamline your workflow, it must be able to understand and interpret medical documentation correctly. This falls under natural language processing (NLP).

Definition: NLP is an AI branch where the technology reads, understands, and makes sense of human language.

Using the data mapping example from above, an effective AI system’s NLP algorithm should be able to decipher and interpret medical documentation accurately. This means the technology could review a medical record with the term “operative procedure” and know that the information should fall under the “procedure description” data field.

Of course, AI developers will also build the system’s knowledge base to ensure the NLP effectively connects the dots of terms that don’t match up character for character but mean the same thing. Developers and users can also add in subject matter expertise experience, other medical terms and abbreviations, as well as lay terms to expand the NLP algorithm’s capabilities.

“Humans can easily interpret the differences, but a machine’s not always going to be prepared for that because it’ll have been trained on a certain dataset. That expectation of knowing the dataset that I give it is specific to me, and I need to customize the solution to understand those points,” Miller said.

Is the AI Complying With Coding Guidelines?

Human medical coders have to know ICD-10-CM guidelines, CPT^® guidelines, CMS reporting requirements, and individual payer reporting requirements. They know that Blue Cross Blue Shield may prefer codes reported one way, while specific states want the codes reported the complete opposite.

The AI coding technologies are developed with standard coding guidelines, which means your diagnosis coding, procedure coding, and HCPCS Level II coding requirements are built in. However, when it comes to generating claims for specific payers, you’ll need to instruct the system on what is needed.

“If you’ve got something specific for your payer and you expect the AI solution to give you an answer based on that, then you have to train it and the system has to know that information,” Miller added.

Check in next month where we’ll examine the remaining factors, such as machine learning capabilities and scaling the technology to fit your organization’s needs.