When a team of emergency physicians at Mayo Clinic in Rochester, Minnesota, decided to take on a research piece on artificial intelligence (AI) for medical billing, they first pieced together a team to get it done.
Jacob Morey, MD, MBA, knew they needed an AI expert, and Dr. Morey found Derick Jones, MD, an emergency physician trained in data and analytics. Medical billing’s impact on finances begged for an expert in that field. Richard Winters, MD, FACEP, is an emergency physician.
The result is a research piece published online in September in Annals of Emergency Medicine that looked at 321,893 adult ED encounters from their health system from January to September 2023. They developed an ensemble model, using natural language processing and machine learning techniques to predict billing codes from clinical notes combined with clinical characteristics and orders. By the end, the researchers said their conclusion matched the theory.
Machines could learn patterns and predict evaluation and management professional billing codes from Levels 2-5.
“It was a validation,” Dr. Jones said. “What we expected the model to be learning was proven in the results. It was aligned and very reassuring.”
While the Mayo team closes the book on this research article, experts in the reimbursement and coding field say the story of AI for medical billing is barely into Chapter 1. One plot could focus on enormous potential to save time and money and streamline a tedious, necessary process of managing reimbursement.
The other storyline warns about putting too much stock in AI too early.
But investigating ways to make things more efficient is worth the work and exactly why the research team said they tackled this research, and the subsequent piece, titled “Artificial Intelligence to Predict Billing Code Levels of Emergency Department Encounters.”
They wanted to shed light on the possibilities that exist to improve how physicians are reimbursed for their services.
“Taking our notes and putting them into billing codes is a repetitive process,” Dr. Morey said. “That’s something that has high potential to be automated by AI. And now, we’ve seen better AI tools over the past few years. Now that technology is available to take our notes and use natural language processing to be able to take that data, make it into more structured data, in a model and automate it.”
The AI used in the study analyzed various factors, such as the number of medical orders, discharge disposition, and specific notes within the clinical record.
Model performance for professional billing code levels of 4 and 5 were AUC-ROC 0.94 and 0.95, accuracy 0.80 and 0.92, and F1-score 0.79 and 0.91, respectively. At a 95 percent decision boundary threshold, Level 5 predicted charts had a positive predictive value of 0.99 and sensitivity of 0.57.
The research team says those numbers tell a story every physician can get behind; AI can save time and money. According to research published in the Journal of the American Medical Association in 2021, administrative expenses are a major proportion and growing source of health care costs in the United States, estimating that they contribute 15-20 percent of total national health care expenditures. In the emergency department, it is estimated that the total cost of these activities (e.g. registration and reregistration, physician time, billing, and overhead) is $61.54 per encounter, including professional billing costs of $38.88 (25.2% of professional revenue) for discharged patients, and 32 minutes of total processing time.
“From the finance perspective, it’s no different than anything else—we want to prove the concept,” Dr. Winters said. “And as we’re thinking about, how do we implement it? There are a lot of other things to consider. We want to make sure that we’re compliant with the regulations that we’re able to adapt. We can’t be assigning improper codes because that would have a big effect on the patients and their finances and organizational finances. Like any other organization thinking about AI, we have identified a space where AI can be helpful. Now, what needs to be built around that to make sure that it is valid and that it’s something we can rely upon.”
Recent advances in AI for medical billing are only possible now because of what Edward R. Gaines III, JD, CCP, calls a generational change in coding standards in 2023. It opened the gates to machine learning, said Mr. Gaines, the Vice President of Regulatory Affairs and Industry Liaison at Zotec Partners, LLC. He has been a member of ACEP’s Reimbursement Committee since 2015 and is a longtime faculty member at ACEP’s Reimbursement and Coding Conference as well as an honorary ACEP member.
Mr. Gaines said the new standards took away a lot of the subjectivity, the old way of coding. Zotec responded by hiring engineers from Google to develop its machine learning system, which they currently use for hospital clients and physician groups of all sizes.
“A substantial portion of our coding is done with that engine, and it allows us to give feedback to doctors at a level we’ve never been able to, so accuracy is a huge part of this,” Mr. Gaines said. “The machine will call out if you’re down coded from a 5 to a 4. It will call out why that happened and will show that maybe you didn’t include certain notes that you should have. The machine can develop a profile about how you can document. That activity would take thousands of hours. The machine can do it with relative ease.”
The AI in market for billing and coding already stands at $2.4 billion and is expected to grow to $8.4 billion by 2033, according to an article published in March at media.market.us. This expected growth doesn’t surprise the Mayo research team, and an important piece to the growth could be patient satisfaction. With less subjectivity built into medical billing, patients could have more information about their
care—particularly related to how much it’s going to cost.
Patients “always ask, ‘What are we going to do, and how much is it going to cost?’ Dr. Morey said. “And we never have a good answer for it because we don’t know how much it’s going to cost until later. And they don’t find out until a month after that. If we can improve the ability to use AI to code these charts, maybe we can also know what it’s going to cost the. That might be a long way off, but it’s a possibility.”
While the study demonstrated the potential of AI in automating the billing process, the authors also noted limitations. The model was developed and tested using data from a single health system, which may not generalize to other health care environments with different coding practices or patient populations.
Additionally, the AI model is currently limited to professional billing codes for ED encounters. Future research could explore the application of AI to other areas of the revenue cycle, such as facility charges, procedural charges, or inpatient billing.
Dr. Morey emphasized that the research represents a “proof of concept,” but much work remains before AI-driven billing can be widely implemented. Continuous monitoring and retraining of the AI model will also be crucial to its success. As coding guidelines change and new clinical practices emerge, AI will need to adapt to ensure continued accuracy.
It still comes down to the physician, said B. Bryan Graham, DO, FACEP, a frequent speaker at ACEP’s Reimbursement and Coding Conference. Graham is an emergency physician at the Cleveland Clinic and Medical Director of the Cleveland Clinic Virtual Emergency Medicine Program. He leads many of the reimbursement-and denial-related initiatives for the Emergency Services Institute at the Cleveland Clinic.
Dr. Graham points out that a lot of coding still comes down to medical decision-making.
“The focus on medical decision making and also the advancements of electronic medical records have allowed there to be an opportunity for AI,” Dr. Graham said. “Frankly, if there’s not good documentation in the charts, it expands on the differential. What was the thought process the physician was going through when they were diagnosing and treating the patient? I think there are still limitations to AI, but as it continues to be perfected and advance—and the electronic medical records continue to become smarter and more customized to how we interact and document—it will improve over time.”
Darrin Scheid is communications director at ACEP.
The post Artificial Intelligence for Medical Billing appeared first on ACEP Now.