What Radiology Tells Us About the Future of AI in Medicine
I attended the 2022 annual meeting of the Radiology Society of North America (RSNA). Over four days in December 2022, nearly 35,000 attendees presented the latest information on all kinds of topics. My impression was AI was prominently discussed and presented just about everywhere. Here are my key take-aways and what each might indicate for the broader industry.
No other medical specialty has adopted AI greater or faster than radiology. According to a survey of 1,000 physicians working at Brigham and Women's Hospital in Boston, Massachusetts in 2001, 64% of surveyed radiologists reported using AI (1).
Number of physicians reporting they use artificial intelligence in their medical practice at Brigham and Women’s Hospital (n=1000)
That survey is certainly not meant to represent the state of the industry, but given Brigham is known for innovation it might be a hint of the future. Consider that the number of AI tools cleared by the FDA for clinical use right now in the U.S. are dominated by radiology applications (2). One begins to see and recognize that this specialty probably has already learned some important lessons.
Distribution of FDA cleared artificial intelligence products in the U.S. by medical specialty. Source: FDA(2)
Key Take-aways from RSNA 2022
Physicians are enthusiastic about the POTENTIAL for AI in medicine, however they remain concerned about safety and quality. Most of the AI tools on the market today are deterministic models (such as those that recognize lesions in a radiograph), thus are trained on large data sets originating from patient populations in the northeastern and west coast of the US. The amount of bias inherent in the AI's on the market are unknown, and represent a new kind of risk for healthcare providers. Key take-aways are:
There is a need to validate local population data sets before implementing an AI model built elsewhere. However, most healthcare delivery organizations (HDOs) do not have data scientists who can shepherd that kind of study. The industry must acknowledge this gap, with AI vendors including such validation as part of their service offering for implementation. This validation should be performed in a way that demonstrates neutrality by the vendor.
Today, deterministic AI tools are marketed by technical factors such as touting their degree of specificity and sensitivity. Clinicians aren't too interested in theory, what they want to see is clinical studies that demonstrate results.
Radiologists are acknowledging their own alarming bias to trust AI more than their own medical training. This can lead to errors and safety issues since AI results are not perfect. Even in radiology it is still early days regarding understanding biases and limitations, both on the AI side and the human using the AI. The industry seems far ahead of the clinicians here on AI acceptance by focusing on use cases and benefits and ignoring the complex reality of how these tools need to be governed.
Using AI tools for diagnostic assistance is the most common use case in radiology today, such as through computer vision recognizing a lesion on in a large series of CT images. Helping the radiologist be more productive and possibly lowering stress is a good thing, but is not seen as driving enough value to drive HDO investments. Other key take-aways are:
More value will be realized when AI assists with treatment decisions, based on a complex set of inputs like patient history, genetic makeup, and many more factors. In the case of oncology, it gets more complicated by knowing what type of cancer, where it is located, and its rate of growth for example. Some vendors and thought leaders are already speaking about how AI will help enable a new era of decision intelligence (3).
AI tools have enormous potential to relieve the burden and toil in everyday work streams, which would generate a lot of value given the current workforce crisis at HDOs. AI-backed probabilistic language models that auto-suggest natural language summaries of clinical findings would be a game changer.
Some vendors are already marketing AI built into their imaging machines, so that a less-skilled technician can perform advanced imaging studies. This same concept should be applied across the healthcare enterprise to increase quality while reducing reliance on human expertise. This will help reduce costs and potentially solve workforce shortage issues for highly trained technicians.
Radiology vendors seem to be coalescing around three business models for AI.
The first was discussed already, which is selling an AI with a deterministic algorithm used assisting clinicians make diagnosis or interpret information, with the future being decision intelligence. Example vendors in this space are AIDOC and Lunit, and new startups like Cubismi.
The second model is around platforms that are used to build, test, host, manage, and govern AI usage across the enterprise. Nuance and Rad.ai are examples of this approach. The value here is all about making AI's manageable in an era expected to be full of them, so that an HDO can select the best of breed clinical AI and manage them centrally.
The third business model is around applying probabilistic AI to make workflows more efficient and reduce toil for care givers. These vendors promise higher asset utilization, greater image quality, better patient experience, and improved care provider work experience. Example vendors in this space are Viz.ai, V7Labs, LeanTaaS, and Microsoft/Nuance with Dragon Ambient Experience (DAX) product.
I expect to see new AI tools targeted at other medical specialties evolve like radiology. Hopefully these companies will look to radiology to learn from its experience. In fact, the RSNA and American College of Radiology have several opportunities to get involved in standard setting. I encourage readers to seek these groups out to learn more.
References:
The Use of Artificial Intelligence in Medicine: A Survey of Physicians. Authors: James G. Anderson, MD, MS; David A. Asch, MD, MS; and Atul Gawande, MD, Published: JAMA Internal Medicine, March 1, 2022
FDA Announces More Than 500 Market-Cleared Artificial Intelligence (AI) Medical Algorithms Available in the United States, Published: February 24, 2023, FDA
Decision Intelligence in Healthcare: A Primer. Authors: David A. Asch, MD, MS; Atul Gawande, MD; and James G. Anderson, MD, MS, Published: JAMA Internal Medicine, March 1, 2022
