Revenue cycle management performance has never been more important. And recent advances in technology, particularly artificial intelligence, offer much potential for healthcare's administrative functions.

The RCM function could lay the foundation to harness technology to contribute to better hospital and health system performance, said Jay Aslam, cofounder and chief data scientist at CodaMetrix. Aslam was part of the team that developed Massachusetts General Brigham's original medical coding AI system in 2016 and has an insider's perspective on the role AI is playing in driving impact in RCM today.

We interviewed Aslam, who has more than 30 years of experience developing AI, machine learning and natural language processing technologies, to talk about his Mass General Brigham AI effort that spun off to become CodaMetrix, his views on the role generative AI can play in RCM, and what he thinks the next five to 10 years look like in healthcare for AI.

Q. You helped create Mass General Brigham's original medical coding AI, which created the spinoff CodaMetrix, your company today. Please tell the story of your AI efforts at Mass General Brigham, what the AI does, and how the spinoff happened.

A. The origins of the founding of CodaMetrix in 2019 began 10 years earlier in 2009 when I signed on as a consultant to a company (VOBA Solutions) working with the Massachusetts General Physicians Organization (MGPO), part of what is now Mass General Brigham. VOBA developed custom systems and performed systems integration for various revenue cycle functions at Mass General, including medical coding.

As is true at most health systems, the burden of medical coding often falls on the physicians themselves (for example, for the CPT or procedure codes) and/or professional medical coders (often for the ICD or diagnosis codes), and the MGPO was particularly keen to relieve the burden of coding from physicians but also to improve the efficiency of their professional medical coding staff.

VOBA and the MGPO knew they had a wealth of data to make their systems "intelligent," but they didn't have the expertise to do so.

I was brought on as a consultant given my expertise in AI, natural language processing, machine learning and statistics, and given the fact I had worked with a VOBA member in the past.

To relieve the physician coding burden, we began by building an AI-based system that could whittle the universe of CPT codes down to just a handful of likely codes a physician needs to consider when faced with a medical coding task.

Essentially, we could learn from historical billing data that, for example, a knee and shoulder surgeon performing a surgery with a given scheduling description would, with high probability, have performed one or more of just a handful of procedures – and we could present a list of the CPTs corresponding to those most-likely procedures, together with their descriptions, for the surgeon to use as a starting point in their coding effort.

The AI-based system learned continuously over time, and given sufficient data, it could learn to tailor its results to a particular surgeon (in our example), vastly limiting the space of most likely codes for a physician to review. This greatly reduced the burden on physicians when they were faced with medical coding tasks. This system was deployed at Mass General Brigham in 2010, and it has been in use ever since – continuously learning.

In that system, we relied on the physician – who knew what procedure(s) they performed – to ultimately choose the appropriate CPT code, but we gained efficiency by providing the physician with a good starting point and the right information to easily perform this task.

If we instead relied on the clinical note, then we could potentially eliminate the involvement of physicians altogether for CPT coding and/or professional medical coders for CPT and ICD coding by predicting codes directly from the clinical note itself.

Such an AI-based system would need to learn the patterns of words and phrases in a clinical note that correspond to any given CPT or ICD code, together with the myriad and varying coding rules dictated by various governing bodies and payers.

Furthermore, if the AI-based system could accurately self-assess its confidence in those predictions, it could perform autonomous medical coding – sending cases direct-to-bill without human intervention when such cases, based on the AI’s self-assessed confidence, guaranteed a specified level of accuracy, while sending the remaining cases, together with the AI’s predictions, for human review.

We developed just such a system and deployed it at Mass General Brigham in 2015, where it has been running successfully and continuously learning ever since – automating medical coding, relieving physician burden, and increasing the efficiency of the Mass General Brigham professional coding staff.

Given the success of this in-house developed and deployed system, Mass General Brigham eventually decided to explore the viability of this technology in the greater healthcare market. Once it was determined this technology could be used and useful well outside the confines of Mass General Brigham, it was decided to spin out a company dedicated to developing and deploying this technology for the greater healthcare industry. Thus, CodaMetrix was born in 2019.

Q. Today you're big on incorporating generative AI into the administrative functions of revenue cycle management. Please describe your vision.

A. Our vision is to increase efficiencies and reduce costs in the U.S. healthcare system; to relieve physician and medical coder burden; and to provide autonomous medical coding with the accuracy and clinical specificity necessary for fee-for-service care, value-based care, population health and beyond. Let me describe each in turn.

First, estimates vary, but administrative and revenue cycle functions account for approximately 20-25% of U.S. healthcare spending – dollars that could be spent on patient care instead – and medical coding is the most expensive component of revenue cycle. Our vision is to apply AI to increase efficiencies and reduce cost in the U.S. healthcare system, starting with autonomous medical coding.

But those same AI techniques can yield insights and solutions well beyond just autonomous medical coding; those techniques and the analysis of their results can also be used to optimize the routing of cases needing manual review to the most appropriate medical coders, identify opportunities for clinical documentation improvement, and pave the way for payer-certified coding algorithms, auto-adjudication, automated pre-authorization, and beyond – all driving efficiencies and reducing costs in the healthcare industry.

Second, our goal is to employ AI to reduce physician burden and allow professional medical coders to operate at the top of their licensure. For the former, let me begin with two anecdotes. My father was a practicing physician until his retirement about a dozen years ago. I remember as a child in the 1970s that my father would make house calls – and I would occasionally tag along – because he had the time to do so and could provide that level of care.

However, by the time my father retired from private practice, he was spending many hours each day on the paperwork needed for reimbursement, pre-authorization and the like – and he was not alone in being subjected to this ever-increasing physician burden that reduces time with patients and drives physician burnout.

Second, I have a relative who recently went through a residency and internship program for radiology at one of the most prestigious medical institutions in the U.S. He told me the story of how the residents would draw straws each week to see who would perform the medical coding for all the radiology cases that week while the others could focus their time on – learning radiology.

Our vision is to employ AI to relieve physician burden and allow physicians to learn and practice their craft.

Even for professional medical coders whose job it is to perform medical coding, the medical coding task can be tedious. Routine cases such as chest X-rays or screening mammograms with no findings do not require the significant skills learned by professional medical coders, and our aim is to automate all such cases – and more – to allow these professionals to operate at their highest level.

Finally, medical coding is the language used to abstract and describe patient encounters, for reimbursement and beyond. At present in a fee-for-service use case, the medical coding need only meet a lower "medical necessity" standard wherein clinically comprehensive coding is unwarranted and often unwanted.

However, for value-based care, population health, clinical trials, longitudinal analyses and more, there is a great need for far more accurate and comprehensive coding, and our vision is to employ AI to provide that level of coding, accurately and efficiently.

Q. What do the next five to 10 years look like in healthcare for artificial intelligence, machine learning and natural language processing?

A. First, a general comment. In the future, I think the AI revolution will be viewed much like the smartphone revolution in the sense that AI will be viewed as a universal and indispensable tool that improves our daily lives, but one we have to learn to use wisely.

Consider your smartphone and think about how much of your daily life – mostly for the better, but sometimes for worse – revolves around this indispensable device. AI will be like that – both universal and indispensable – and it is up to us to learn to leverage the benefits while minimizing the costs.

Within healthcare, autonomous medical coding is just one application of AI. And while just a handful of years ago, autonomous medical coding was viewed as the province of large academic medical centers who could afford to experiment with cutting-edge technology, it is rapidly being viewed as a necessary and indispensable tool needed by all health systems – in much the same way the original smartphones were once viewed as cutting-edge technology for early adopters but rapidly became indispensable tools for everyone.

AI will be like that for all aspects of healthcare including diagnostics, treatment planning, drug discovery and design – virtually everything. The combination of vast amounts of data, computational resources and the latest AI algorithms will enable rapid improvements in all these areas, and we are seeing such improvements today.

And my parting comment and vision for the future is that AI will not entirely replace human effort but rather augment humans, and that human-in-the-loop, AI-augmented systems can achieve results better than AI or humans alone. AI is a powerful tool that can and will be used by, for and alongside humans in the healthcare industry to drive efficiency and achieve performance.

Follow Bill's HIT coverage on LinkedIn: Bill Siwicki
Email him: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication.