Revisited: FDA's AI Medical Device Approvals

About a year ago, I dug into FDA’s newly-released list of its AI medical device approvals. In that post, I noted some surprising findings: AI approvals were (1) becoming more concentrated, with a small group of companies winning a large percentage of the approvals, and (2) most approvals were going to established medical device titans, like Siemens and GE, rather than to startups. Both findings showing that AI in medicine was providing sustaining innovation for established companies more than disrupting the established ecosystem.

A few weeks ago, FDA provided new data:

October 19, 2023 update: 171 Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices were added to the list below. Of those newly added to the list, 155 are devices with final decision dates between August 1, 2022, and July 30, 2023, and 16 are devices from prior periods identified through a refinement of methods used to generate this list. 

With this release, I wanted to see if my earlier findings still held, how fast innovation was accelerating, how quickly were more companies getting involved — and what other surprises lay within the data.

Anti-exponential: medical AI innovation is not accelerating

Scoring points: approvals by year

Probably most interest of all findings is that AI innovation in medicine — at least by the measure of AI device approvals — isn’t accelerating at all! The graph below shows good, steady growth in approvals per year (an average 26% increase per year over the last 4 complete years) but not acceleration:

Of course the number of approvals depends on at least two variables: how many applications are submitted, and how many are approved. It’s possible that the rate of invention (i.e. the creation of new algorithms and tools) is accelerating but that the FDA approval process is the rate limiter.

Joining the team: new companies per year

In terms of broadening the group of innovators, the list this year includes 19 entities that previously had not appeared, including stalwarts of diagnostics like Beckman Coulter and Biomerieux:

• AgaMatrix Inc.

• AlgoMedica

• Appian Medical Inc.

• Beckman Coulter, Inc.

• Biomerieux, Inc.

• Bruker Daltonics, Inc.

• ClearView Diagnostics Inc.

• Cydar Ltd.

• Gauss Surgical Inc.

• IRIS Intelligent Retinal Imaging Systems, LLC

• LabStyle Innovations Ltd.

• Matakina Technology Ltd.

• Monarch Medical Technologies

• Pathwork Diagnostics Inc.

• QView Medical, Inc.

• Renalytix AI, Inc.

• Stratoscientific, Inc.

• Tyto Care Ltd.

• Zepmed, LLC

Liked approvals per year, this is an increase but not an acceleration. As shown in this graph:

So neither the number of approvals nor the number of companies winning those approvals is accelerating.

Less lonely at the top

Last year I noticed that approvals were becoming more concentrated — with only 18% of companies on the list gaining more than half of total approvals (this became 20% with revision of the list by the FDA; see notes at bottom). So far in 2023, you’ve got to include 30% of total companies to get to 50% of total approvals — so approvals are going to a broader group. Hard to say from the data whether this is a trend or just variation year to year, and I’ll plan to revisit when next year’s data comes out.

Your grandfather’s medical AI?

Age of the players

I was surprised last year to find that the average age of the top 6 companies in the approvals list (including all years) was 89 years!

The most surprising point hidden in the dataset was that a lot of the action in medical AI isn’t coming from Big Tech companies known for AI like Apple or Amazon or Google/Verily but from much much older Big Med Tech industrial companies, with GE (founded 1892) and Siemens (founded 1847) taking the number 1 and 2 spots, respectively.

That has changed a bit. The top 12 companies for 2022-23 have an average age of 58 years - so we can count that as a major shift towards new players like Annalise-AI (4 years old), Viz.ai (7 years old), Hyperfine (9 years old), and Aidoc (7 years old).

This is also reflected in the continued drop in the median age of companies receiving at least one approval, as shown in the graph below.

The bump in average age for 2023? That’s in part from the addition to the list of Cedars-Sinai Medical Center (founded 1902) and MD Anderson Cancer Center (founded 1941). Will such venerable institutions continue to move onto the list? Given the accumulated medical knowledge there, let’s hope so.

The graph above just looks at the age of companies “in the game,” with at least one approval. But that means that for 2022, for example, it counts GE, with 17 approvals, the same as United Imaging — with just one.

The graph below gives us more insight into that game, showing what percentage of players are “older” (>= 20 years) versus younger.

Points scored by young vs old

It’s also important to know whether older vs younger players are racking up points (i.e. approvals). And here we see that despite the influx of younger companies the older companies are more than holding their own. In fact, as noted last year, the percentage of approvals going to older companies is steady or possibly increasing.

Big tech MIA

Note that although older companies are holding their own, those companies mostly aren’t Big Tech. Since 2017, Apple’s only gotten 2 approvals, Google’s Verily 2, Microsoft 1, and Amazon 0. That’s a total of 5 out of 531 approvals: less than 1%.

What’s it all mean

At least on the medical side, as measured by FDA AI device approvals, AI innovation is chugging along — but not accelerating. This comes as a bit of a useful corrective to the relentless AI hype. It also raises the question as to whether the rate-limiting factor here is the FDA approval process itself (not hard to believe considering the slow pace of drug approvals), or something intrinsic to the market.

More companies are entering the space and getting approvals. Most of these are younger, but the entrance of MD Anderson and Cedars-Sinai stands as an example for the well-established hospital systems; an example which I hope will spur more participation from the healthcare establishment not just in adopting AI innovation but in helping to create it.

The percentage of approvals going to the older companies is, if anything, increasing. This is classic sustaining innovation in the medical device space, as companies like GE add AI components to improve their existing product line. It remains to be seen who in the space will start to significantly outpace the innovation at those older companies and seriously disrupt the space.

Finally, despite great stated and demonstrated interest in health, Big Tech is still not a major presence on FDA’s list. Given the health moves made by Apple, Google, Amazon, Microsoft and others, this is a little surprising — especially on the Apple side, given their focus on device innovation. Will this change, whether through in-house innovation or acquisition? Time will tell.

---

Notes

1. I’m definitely not an FDA expert, just a doctor and tech guy trying to figure out what’s going on in the world. If you’ve got special insight into any of this, or corrections to the above, please let me know.

2. The FDA’s AI list appears to be incomplete, leaving out at least 4 de novo approvals utilizing AI that I was able to find by cross-checking other FDA references: one from 23andMe, one from Apple, one from Renalytix AI, and one from Viz.ai. Those 4 are in FDA’s searchable approvals database, but not in their AI list. I’ve included those four in the analyses above.

3. The FDA’S list includes this note: “Of those newly added to the list, 155 are devices with final decision dates between August 1, 2022, and July 30, 2023, and 16 are devices from prior periods identified through a refinement of methods used to generate this list.”

4. I was unable to find out when these two companies were founded, and so they’ve been left out of age calculations. If you know, please share!

   • Yukun (Beijing) Technology Co., Ltd, filer of item K213986 for CerebralGo Plus

   • AssembleCircle Corp., filer of item K220903 for WebCeph

Disruption for Doctors 2: Healthcare Examples

Recap of part I: disruptors are worse, not better

In my last post, I introduced the concepts of innovation developed by Clayton Christensen of Harvard. In Christensen’s paradigm, “sustaining” innovation enables incumbent organizations to better please their existing customers, while “disruptive” innovation weakens incumbents by creating new markets with cheaper but arguably worse technology.

As an example of “worse” disruptive technology, I talked about Netflix, which originally sent customers DVDs by postal mail. That seemed ridiculous to many who liked being able to just walk into a Blockbuster video store and then walk out with a video in a short time.

And the incumbents at Blockbuster apparently laughed and laughed . . . but Netflix was able to siphon off Blockbuster customers who were more price-sensitive. Then, when Netflix adopted streaming technology, they took all the remaining customers because at that point they were both cheaper AND faster.

But healthcare is not video streaming, so what about some examples of disruption in healthcare?

Non-IT-based healthcare disruption

There has been a lot of disruption within healthcare in the last decades; with some of it mostly related to information technology (IT) and some not. Let’s look first at some of the non-IT stuff.

Some examples:

• doctors getting disrupted by ancillary providers like nurse practitioners, who have many fewer years of expensive training, and are paid less

• primary care clinics and emergency departments getting disrupted by urgent care centers with a limited menu of services, and without the legacy costs of giant hospital-centric systems

• specialists getting disrupted by other specialists (see next section, below)

Note that, as with our Netflix example, the disrupters are arguably worse in certain ways than the doctors they are trying to replace (sorry, “augment” 🤣). It’s hard to argue, for example, that nurse practitioners are better than primary care doctors. The important thing, however, is that they are good enough for some section of the patient population — and that they are cheaper.

Also, although these disruptive innovations aren’t mostly related to IT, they are dependent on the IT innovations of the last 40 years: it wouldn’t be so easy to run an efficient urgent care center without inexpensive personal computers, etc.

A simple menu means costs can come down.

Example of non-IT healthcare disruption: heart disease treatment

Back in 2000, in an article entitled “Will Disruptive Innovations Cure Health Care?” Christensen himself provided an example of healthcare disruption. He and his coauthors talk about coronary angioplasty (or PCI, for “percutaneous intervention”), which enabled cardiologists to treat heart disease patients by inserting tiny tubes called stents into blood vessels — instead of requiring cardiac surgeons to perform cardiac bypass grafts (CABG). This is disruptive technology, for sure, and it’s a good illustration that disruptive technology doesn’t have to be information technology.

PCI didn’t eliminate heart surgeons, of course, but it did drastically reduce their activities: between 2001 and 2008, the annual rate of CABG in the US went down by 30%! So the incumbents (cardiac surgeons) lost a lot of revenue to the disruptive practitioners (interventional cardiologists) doing the PCI.

Keep in mind that CABG operations were, and are, better in certain circumstances but the disruptive PCI was generally cheaper because cardiologists are cheaper than cardiac surgeons, and PCI can often be outpatient. So the general idea of “disruption by worse, but cheaper” holds for this example.

The latest disruptor?

Interestingly, more recent data to 2016 demonstrated a roughly 40% decrease in both CABG and PCI — possibly due to more effective medicines (like statins and beta-blockers) reducing the need for surgery. So perhaps medical treatment for heart disease is now disrupting PCI, which previously disrupted CABG? Which might shift revenue back to primary care providers, whether doctors or nurse practitioners or other.

IT-based healthcare disruption

Disruptor of many things

Disruptor of many things

The reason that nowadays we focus on IT when we think about disruption is that for about 40 years, most business disruption in the world has been driven by the IT revolution (further reading: Why Software is Eating the World).

Think about the efficiencies and cost savings made possible by the PC, the internet, cloud computing, and the mobile phone.

Now think about the fact that healthcare has really only adopted one of those innovations: personal computers. Ouch.

Which means that companies that are very, very good at IT are turning their attention to healthcare to see if they can begin picking off the low-hanging fruit from set-in-their-ways old-school health systems. And this includes those using the very latest technology: artificial intelligence (AI).

Some examples:

• efficient lower-overhead systems (like Walmart Health) disrupting hospital-based systems (AND the urgent care companies!) with lower-cost systems that provide only the most commonly-needed services, and that can leverage their enormous portfolio of existing stores for office space. Also remember that Walmart got to be Walmart by being really good at IT (for supply chain management): we can expect efficiencies that most hospital systems can’t approach.

• deep-pocketed tech companies, like Amazon Care, disrupting primary care by using tech to run clinics at lower cost. How can Amazon be successful? By using the same online tools they use as the biggest online retailer in the world and, less well known to the public, as one of the leading providers of cloud computing services to other tech companies, via their AWS division. And they are on a healthcare spending spree lately, having just purchased One Medical [Note: the day after publishing this, Amazon announced they’re shutting down Amazon Care and focusing on One Medical].

• AI-based tools pulling some services away from expensive specialists, with one example being IDx-DR (see next section, below).

These IT-based disruptors are trying to use IT to create simpler, easier to use, cheaper versions of existing healthcare services.

Example of AI disruption: IDx-DR

Digital Diagnostics is a company based in Coralville, Iowa, far from tech centers like Silicon Valley or Boston. Using artificial intelligence computer vision techniques, they’ve built IDx-DR: the “first and only FDA authorized AI system for the autonomous detection of diabetic retinopathy.”

Translated, that means they have a machine that can take pictures of your retina and determine if you have diabetic retinopathy (DR, a debilitating problem that often occurs in poorly controlled diabetes).

Before IDx-DR, the only way to diagnose DR was for a very highly-trained and highly-paid ophthalmologist to look at your eyes. With Idx-DR, the diagnosis is automated and requires no input from any doctor, much less an expensive specialist.

This is the first machine capable of making a medical diagnosis by itself, and according to the website is “now part of the American Diabetes Association’s Standard of Diabetes care”.

Interestingly, it is marketed as a tool for primary care clinics, who can now avoid sending diabetic patients to expensive ophthalmologists for this diagnosis. That is a boon to primary care docs (and their patients) — but with 37 million diabetics in the US alone there is no real technological reason why this shouldn’t eventually find its way to a drugstore near you, right next to that automated blood pressure cuff.

Used to require a doctor

Also used to require a doctor

Example of AI disruption: ChestLink

Another great example of AI-based disruptive innovation has been developed by Oxipit. Their ChestLink app, which is approved by European regulators but not yet the FDA, uses AI to read chest x-rays. It is apparently sensitive enough to identify the presence of problem, but not (yet) as good as a human radiologist to identify what the specific problem is.

ChestLink’s main use, as a result, is to automate the reading of the up-to-80% of chest x-rays that are normal, freeing up radiologists to focus on those images flagged by the software as suspicious. Note again that the software isn’t necessarily as good as a radiologist in identifying a specific problem, but it is good enough for a very common task that takes up a lot of human radiologist time.

How is this disruptive? Well, because using the software to verify x-rays as normal is so much cheaper than using a radiologist for the same purpose:

1. rich countries with radiologists won’t need as many, so the cost of this one small facet of healthcare (i.e. reading chest x-rays) will go down.

2. in poor countries, the software isn’t competing against radiologists, it is — in Christensen’s terms — competing against “non-consumption.” It can thus create a new market for the use of chest x-rays, bringing that beneficial medical technology to many more people.

From https://oxipit.ai

From specialists to generalists, from physicians to nurse practitioners, from hospitals to supermarkets

In Christensen’s 2000 article (echoed later in his 2016 book The Innovator’s Prescription) Christensen presciently wrote about many of the now-established trends in healthcare:

We need diagnostic and therapeutic advances that allow nurse practitioners to treat diseases that used to require a physician’s care, for example, or primary care physicians to treat conditions that used to require specialists. Similarly, we need innovations that enable procedures to be done in less expensive, more convenient settings—for doctors to provide services in their offices that used to be done during a hospital stay, for example.

All of this, of course, is commonplace in the healthcare of 2022: urgent care centers staffed by nurse practitioners, outpatient surgicenters, more techs working throughout healthcare.

But Christensen wrote that in the world of 2000: Google was 2 years old, AI was still frozen in a decades-old “winter”, and the iPhone existed only in the mind of Steve Jobs. The advent of two decades of consumer technology has not only accelerated the healthcare changes that Christensen envisioned, but also accelerated something he doesn’t seem to have thought much about: selfcare. That is, the ability of the consumer to treat their own illness, and to maintain their own health.

And in my next post, I’ll talk about healthcare moving not from doctors to nurse practitioners, but from nurse practitioners to . . . us.

Stay tuned.

---

This post draws on themes from my keynote talks.