The Economics of Machine Learning

How will emerging technology affect radiology in the near future?economics of machine learning

"DeepRadiology Announces the World’s First Fully Autonomous Radiology Interpretation System.” This news story, released during RSNA 2016, was followed by two statements:

 1. “ The system is able to produce final interpretations and reporting on medical imaging studies and reporting on medical images without the need for a radiologist,” and

2. “ The device is … currently being evaluated by the U.S. Food and Drug Administration.”

I do not know if the system described is real, if this is just hype, or if it is somewhere in between. I do know that this announcement illustrates the need for radiology to respect this emerging technology. As Keith J. Dreyer, DO, PhD, FACR, recently articulated during a presentation to the ACR Board of Chancellors: Radiology should take ownership of this technology — before it takes ownership of us. Ownership includes directing the economic actions related to the evolving technology. In this column, I will discuss several short-term (as in, the next three to five years) economic considerations and activities.

One possibility is that the diffusion of this technology will be incremental, such as an algorithm to detect and flag a focal condition (for example, cerebral hemorrhage) for further review. Under this scenario, it is doubtful any specific action will be necessary in regard to the radiology CPT code set. But imagine a scenario in which an “autonomous” system, like the one mentioned above, gains FDA approval. At that point, we should expect that the FDA-approved vendor will pursue a Current Procedural Terminology (CPT)TM code to report the service. The code or codes created at the CPT Editorial Panel will largely depend on how the technology affects the work of providing a radiological interpretation. An early question may be, is there physician work involved at all? For example, will the deep learning interpretation involve radiologists (or other physicians) or replace them altogether? If there is no work, then there is no CPT code. But what if radiologists use the input from the deep learning system to complement their interpretation, resulting in more work?

Under this scenario, the CPT code may be an “add-on,” such as was the case when mammography computer-assisted detection emerged. If the technology inherently alters the work of the typical interpretation of a specific service, a revision of that service’s CPT code may be in order. For example, if bone age studies someday are “typically” interpreted alongside a deep learning application, the CPT code for bone age study interpretation may require a completely new code and descriptor.

If new codes are created, the next step would involve valuation. The base question remains the same: How is a radiologist’s work changed using deep learning? And how does this change in work translate into the currency used in the Medicare Physician Fee Schedule relative value units (RVUs)? And what effects will the technology have on the valuation of the imaging services to which it is applied. For instance, if deep learning provides an interpretation of a head CT, does that somehow lessen the value of the head CT code itself or negate its value altogether? Or is there an increase in the value of the head CT code to capture the increased work of merging the radiologist’s and the deep learning interpretation? Are there related activities involving radiology (for example, providing testing and validation or developing validation protocols)? And what about the cost of the technology for the facilities or physicians’ offices, generally referred to as the technical component (TC)? How will those entities be paid for the hardware and software necessary to provide this technology, and how will this fit into the broader CMS practice expense methodology? What impact could machine learning have on TC-related activities (such as protocols or image acquisition) and will that affect these payments? These are all important questions for us to ponder.

The emergence of deep learning will affect many industries that rely on large data sets to make predictions and interpretations. Radiology is not alone. Therefore, the Commission on Economics is already scenario-planning, as is the rest of the ACR. I have described several short-term economic decision points potentially on the horizon. As the leaders in imaging technology, it is incumbent on the ACR to take ownership of deep learning, making sure the technology is safe and contributes to high-quality patient care.

dr silvaBy Ezequiel Silva III, MD, FACR, Chair

Share this content

Submit to FacebookSubmit to Google PlusSubmit to TwitterSubmit to LinkedIn