In the final post in this series, I’ll discuss the approaches that can make a very necessary and tedious part of administering healthcare more automated and where it can go from here. If you examine the autonomous coding efforts over the last decade or so, it started with using historical billing data that using statistics could narrow down the list of codes to a handful for the provider to choose from. As Jay Aslam mentions in a recent article in Healthcare IT News, a surgeon scheduling a knee 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. This worthwhile initial approach made incremental progress by accelerating the process and making it less tedious since the surgeon did not have to spend too much time looking for a code after spending a good bit of time performing a surgery. This could improve over time with continuous learning if the results are collected and used to further train the model. This might seem like common sense but it creates work and medical centers have shown that they have higher priorities!!
This approach relied on the physician to do the coding ultimately, probably with human coders reviewing the code and the supporting note to make sure it was all done properly before submitting to the insurance company for payment. The evolution of deep learning and its astonishing ability to learn learn the patterns of words and phrases in a clinical note that correspond to any given CPT or ICD code, and combine that with the huge combinations of coding rules dictated by various insurance companies is poised to usher in a new era where everything, including direct submission of the codes to the insurance companies, can be done with AI. At least the Codametrix system, and perhaps others, can self-assess their confidence in the codes that the AI engine generates and decide to submit direclty to the insurance company or flag it for a human to review.
I discuss this in my book, AI Doctor: The Rise of Artificial Intelligence in Healthcare, that the administrative costs of healthcare delivery in the US are 4-8 times higher than other OECD countries. Why? Because we have many insurance companies, each with their own rules and requirements. Each provider sees many patients and a dizzying combination of insurance companies cover these patients. The provider needs to deal with all of this paperwork and meet their requirements to get paid. This leads to the high administrative cost, which contributes to the higher per capita cost of providing healthcare in US than any other country. And, we don’t get better outcomes for it. Our outcomes are the worst of all OECD countries and in many cases lower than the outcomes of middle income countries like Slovenia. As such, the potential benefits of using the emerging AI solutions to lower administrative costs of care will be disproportionately higher in US. This is much needed as the high cost of care not only uses up resources that can be used for more productive activities but also is a huge threat to US government’s fiscal health long-term. Oddly, medical coding-related activities are the highest driver of the administrative costs in healthcare.
While inadequate documentation is the leading cause of billing code denials by payers, AI is well on its way to fix that also. Some of the emerging AI solutions have the capability to provide real-time feedback to the provider about their documentation quality. This means that any issues with the documentation can be addressed on the spot. Also, as the capabilities of AI expands and more data becomes available to train the models, it’s possible to optimize these models for each specific insurance company. This will also work from the payer side. We expect that they will increasingly use AI to review the submitted codes and the documentation and auto-adjudicate the payment decision. Soon, we will even have pre-authorization letter created by the providers based on what has been documented in an encounter note and those letters to be auto-adjudicated by the payers. Also, as AI’s capability in this area matches or exceeds humans’ with reliable consistency, more complex coding to meet the needs of population health management initiatives, value-based care, long-term analysis, and more can be dialed up.
While all of this is much closer than it has ever been, it is important to keep in mind that given the complexity of all of this and the unstructured and fragmented data that are needed, we will see these solutions emerge over time and with varying levels of proficiency. However, for autonomous coding, there are now solutions that are being used in a number of medical centers and the early results are very promising. As such, I expect autonomous coding to be the way most coding will be done in the next few years.