No, the sky is not falling — but it’s worth considering an enormous problem surrounding access to medical care, AI, accurate diagnosis, and why AI may need to be not just equivalent to physicians — but orders of magnitude better.
I was recently reading Eric Topol, MD's newsletter Ground Truths (subscribe if you’re not!) and came across this tidbit from this Dutch AI skin app study:
A similar example of this comes from a large experience with a skin AI app for detection of cancer, from a study of nearly 19,000 people in the Netherlands who used the app compared with matched controls. There was a 32% increase in potential skin cancers detected, but they were mostly determined to be benign. The study was based on claims data (a substantial limitation), and while there was a net of more cancers detected, the cost for detecting one additional pre-malignant skin lesion was very high (over $2700).
I haven't really been able to stop thinking about this, and the ramifications of AI at scale: massively higher rates of referral to dermatology even with a pretty good AI app (sensitivity of 87–95% and specificity of 70–78%). The app was offered to 2.2 million adults to use for free, and almost 19,000 used it at least once. But imagine if everyone used it: I mean, who doesn't have a mole or a growth or a thing on their body that they'd love a medical opinion about? What if 2.2 million users signed up — which will certainly happen as we all become more comfortable and more familiar with AI? Users in this app were about 2.5-4x more likely to have a benign lesion removed in the hospital. Is The Netherlands ready to build 4x more operating rooms? Hire 4x more anesthesiologists? Bring on twice as many dermatologists and GPs to review and remove moles, and twice as many pathologists to look at said moles for cancer?
I won't bore you with too many biostatistics details, but now hopefully you're starting to see why accuracy matters so much more as access to a test (for example, an AI model) increases. Let's look at the tool in the US:
- There's 330 million Americans in the US. Imagine even 10% of them have access to a free AI mole tool on their smartphones, and they take 2 pictures each. 60 million images would generate 150,000 additional tele-dermatology referrals... for our 12,000 dermatologists in the US. (Not to mention the 900,000 additional surgeries for benign lesions per year.)
"But Graham," you say, "Why not just make the tool more accurate?" And to that I'd say "My point exactly; we'll need our tools to have drastically higher accuracy." But the authors wisely looked at this, and to some degree you just start to get into a numbers game, trading off sensitivity (referring all potential cancers and not missing any) for specificity (only referring definitive cancers). Here's their figure — it looks like most things in medicine: no magic solution.
I'd be bereft not to mention the other side of this coin, of course; while I've mostly focused on the additional workups for benign lesions, this tool did discover more malignancies. Absolutely some lives were saved because some melanomas were caught and fully removed. This is a Great Thing. But again, the question ultimately boils down to what cost? I hate to get all actuarial on you, but what's the ratio we should accept? Is stopping 1 fatal melanoma worth a cancer scare and a surgery for 5 other people? Probably. 10? Sure. 1,000? 10,000? A million? What about the societal cost of doing all those extra surgeries — or the opportunity cost of doing those extra surgeries instead of a surgery for someone else who needs it?
And now let's extrapolate this outside of the realm of just moles or dermatology — imagine an AI chatbot on your phone that gives you instant access to a medical resource, 24 hours a day, 7 days a week. Have some weird little ache or pain in your body (which is normal and part of the human experience) — why not just ask Dr. Chatbot? Finger tingling for a few minutes? The Chatbot won't mind giving you some advice. And without good guardrails and good recommendations from Dr. Chatbot (especially in our litigious and extremely-CYA American system), just imagine the additional imaging, blood tests, EKGs, and referrals to specialists and the ER that a Chatbot could create.
(Don't think this would happen? Look at your own smartphone usage. When was the last time you had a question about something — literally anything — and you didn't immeidately pull out your phone to find out the answer?)
We're already seeing this trend to some degree with patient access to secure messages with their doctors — as you reduce barriers to access to care (again, generally a well-intentioned, overall-Good-Thing), you increase the workload on physicians and increase utilization and testing. If we think we're ordering too many CTs today...just imagine the future.
I'll end by calling out the current American system for accessing care: it sucks. No doubt about it. It sucks for the un- and under-insured. It even sucks for the insured. It's no fault of a patient that they're trying to get a question answered about their health or their body. That's a reasonable thing to expect from a healthcare system. And again — I'm cautiously optimistic that AI can really help and can make things better for doctors and for patients. (Insert plug to read our outline for responsible AI in medicine, called The Physicians' Charter for Responsible AI.)
But it's clear to me that this will not be a walk in the park, or a leisurely stroll, or that we'll just be able to say "Let's add AI for our patients" and it will magically happen without careful thought and numerous unintended consequences. And many missteps. AI is going to transform society to the point that it will be unrecognizable. Medicine will be just as impacted, if not more. But if we're not careful, thoughtful, and cautious in our approach, we run the risk of over-utilization, over-diagnosis, and over-workup, leading to unnecessary anxiety and complications for healthy people and worsening access for patients who urgently need our attention and care.
(Want a more academic analysis of the dermatology paper? See the accompanying opinion piece, AI-based skin cancer detection: the balance between access and overutilization.)