Discover how Aidoc’s AI platform offers seamless end-to-end integration into a facility’s existing IT infrastructure enabling implementation of AI at scale.
Discover how Aidoc’s AI platform offers seamless end-to-end integration into a facility’s existing IT infrastructure enabling implementation of AI at scale.
Rad Perspectives: Reflecting on 2019 Healthcare AI Predictions
Table of Contents
In 2019, I co-wrote an article with our CEO Elad Walach for AuntMinnie titled “How AI will become an integral part of standard care.” Well, it’s 2023, so if that article represents our freshman effort, we’re in our senior year now, and it’s time to look back.
How did we spend the past four years? Were our predictions correct? If so, did we act correctly based on those predictions? If not, where were we wrong, and why? And finally, what do we predict for the next four years?
The Pandemic
Before I review what we wrote in 2019, I cannot avoid addressing the tsunami we could not foresee that January–the COVID 19 pandemic. It’s fair to say that the pandemic caused irreversible upheaval in every aspect of human life. This is nowhere more the case than in the world of medicine, and of course this includes the med-tech sector and the medical AI ecosystem within it.
Many medical AI developers, including Aidoc, scrambled to produce COVID-related algorithms to save the world, with varying levels of success. As a market observer, I learned a lot about the folly of hasty algorithm design and deployment. In any event, these products seem to be of declining relevance in 2023– let’s hope it stays that way.
But more importantly, COVID laid bare the brittleness of even the world’s most advanced medical systems in handling major stressors. Staffing shortages were more apparent than ever, and fiscal frailty never more trenchant. Seemingly overnight, entire hospital systems crumbled under the load and had to pivot to a crisis stance which persisted for the next 1-2 years. Nobody had the budget to consider buying AI products that weren’t specifically intended to extract them from the crisis, and VC investment in new endeavors similarly retracted.
We at Aidoc were fortunate and got through relatively unscathed, although certainly challenged and humbled. But the silver lining in this viral cloud was that the crisis confirmed–in a sort of ad extremum manner–that our basic assumptions about the role of AI were correct. In order to give the best care to the most people, medical systems already operate very close to the upper limit of their resources, human and otherwise. The difference between those providers that survive the next big stressor and those that don’t will be technological advances that significantly increase the efficiency of their existing resources.
ROI, Insurance Codes, and Value-Based Reimbursement
Financial considerations were and always will be major concerns for AI adoption in medicine, but in the the post-pandemic recession of 2020, they were foremost. Maybe they still are. In 2019 we discussed the importance of ROI and value-based reimbursement, and those points remain as salient now as they were then.
Thankfully the four intervening years have given us more thorough models to demonstrate the real-world efficacy of medical AI. For example:
A Yale study from 2022 showed that a modest reduction in reporting TAT for patients in the ER with intracranial hemorrhage translated to about a 2-day reduction in hospital length of stay.
As academia and major medical institutions increasingly recognize the benefits of AI, regulatory bodies are following suit. In 2021, the American Medical Association approved the first CPT code for use of artificial intelligence in radiology. In our 2019 article, we hoped they would. We got our wish, sort of; the approved code was designated as class III, which means its purpose is to gather information for research only. Therefore, it is a long way from opening the door to payor reimbursement for use of AI. But, it’s a start. The same year, Medicare awarded its first New Technology Add-on Payment (NTAP) designation for an AI product designed for stroke detection. This obscure congressional provision from 2021 may prove to be very important for medical AI–the expected lag time until class III CPT codes become reimbursable codes is measured in years. I would therefore suspect many more AI applications will receive NTAP status in the coming years.
FDA Support for AI
Ranking AI companies with more than 1 software product cleared by the FDA from 2018 to Oct. 5, 2022 representative of FDA product codes MYN and QFM. Source: Asher Orion Group on LinkedIn
Since 2019, FDA support for AI devices has been tremendous. Despite the pandemic, and despite going through 5 FDA commissioners in this period, the FDA has processed about 350 new clearances for AI-containing medical devices, meaning well over half of all FDA-cleared AI-mediated devices in existence have been processed in the last four years. The distribution of these clearances is also interesting.
Naturally, large multinational corporations like GE and Siemens lead the pack in the number of new FDA-cleared AI products. But even so, their many products typically revolve around scanner function or existing PACS platforms, as shown in the graphic on the left from Asher Orion group, delineating between these and what they call “true” imaging AI products.
As such, the bulk of new clearances are by small companies, each having one or two products. I say this because both interest and competition are high, and the market is being driven by numerous niche designers coming to solve specific problems. This is not only going to be an ongoing challenge for the FDA, but also for the consumers of these products, who will need to find a way to deploy and maintain AI products from numerous vendors, usually on large and sometimes antiquated hospital infrastructure.
Developing an IT Ecosystem
A challenge we discussed in 2019 was the difficulty of hosting a wide array of AI solutions on one platform. This is why we have spent the last four years developing our AI operating system (aiOS™), which is intended to be a vendor-agnostic platform to launch, maintain and integrate a wide variety of AI solutions at enterprise scale. While launching our own products at various partner sites around the world, we noticed that every launch was slightly different. Sometimes we required creative solutions for the varying algorithms to work well together. Having a diverse client base with a wide variety of needs allowed us to gain market-leading experience in what it takes for AI to work in an environment that has never used it before.
Our aiOS™ combines these launch experiences to make it relatively painless to launch AI at scale, and not only with our own products, but with other vendors as well. We hope to be able to provide the output of these various solutions in an orchestrated, digestible format at the point of care across the medical system, from the community clinic exam room to the high-tech downtown operating suite.
Looking back at our 2019 post, I think we were mostly right, and things mostly worked out well. I’m proud that numerous studies validated not only the performance of our products, but the return on investment they can provide. We were taken by surprise by COVID-19, but I’m relieved that we successfully weathered the global pandemic and subsequent recessions; I think that says a lot about our resilience and resolve.
I’m grateful for each and every FDA clearance we have earned on our way to being in the top 5 for AI product clearances. We were hoping for more recognition for AI at the payor level, but we’re sanguine on the prospects of the current CPT codes and NTAP status our sector has achieved.
Finally, we are really excited about the prospect of delivering a huge array of solutions to institutional consumers using our aiOS platform. We are only at the beginning stages of this part of our journey, which means we will probably have to wait another few years before we can reflect on what we’ve accomplished.
