Letting AI Physicians Into the Guild

Let’s be honest: we’re going to have AI physicians. 

Now, that prediction comes with a few caveats. It’s not going to be this year, and maybe not even in this decade. We may not call them “physicians,” but, rather, may think of them as a new category entirely. AI will almost certainly first follow its current path of become assistive technology, for human clinicians and even patients.  We’re going to continue to struggle to fit them into existing regulatory boxes, like clinical decision support software or medical devices, until those boxes prove to be the wrong shape and size for how AI capabilities develop.

But, even given all that, we are going to end up with AI physicians.  They’re going to be capable of listening to patients’ symptoms, of evaluating patient history and clinical indicators, and of both determining likely diagnosis and suggested treatments.  With their robot underlings, or other smart devices, they’ll even be capable of performing many/most of those treatments.

We’re going to wonder how we ever got along without them.

Many people claim to not be ready for this. The Pew Research Center recently found that 60% of Americans would be uncomfortable if their physician even relied on AI for their care, and were  more worried that health care professionals would adopt AI technologies too fast rather than too slow. 

Still, though, two-thirds of the respondents already admit that they’d want AI to be used in their skin cancer screening, and one has to believe that as more people understand the kinds of things AI is already assisting with, much less the things it will soon help with, the more open they’ll be.    

People claim to value the patient-physician relationship, but what we really want is to be healthy.  AI will be able to help us with that.

For the sake of argument, let’s assume you buy my prediction, and focus on the harder question of how we’ll regulate them. I mean, they’re already passing licensing exams.  We’re not going to “send” them to medical school, right?  They’re probably not going to need years of post-medical school internships/ residencies/fellowships like human physicians either. And are we really going to make cloud-based, distributed AI get licensed in every state where they might “see” patients? 

There are some things we will definitely want them to demonstrate, such as:

• Sound knowledge of anatomy and physiology, diseases, and injuries;
• Ability to link symptoms with likely diagnoses;
• Wide ranging knowledge of evidence-based treatments for specific diagnoses;
• Effective patient interaction skills.

We’ll also want to be sure we understand any built-in biases/limitations of the data the AI trained on. E.g., did it include patients of all ages, genders, racial and ethnic backgrounds, and socioeconomic statuses? Are the sources of information on conditions and treatments drawn from just a few medical institutions and/or journals, or a broad range? How able is it to evaluate robust research studies from more questionable ones? 

Credit: BMJ

Many will also argue we’ll need to remove any “black boxes,” so that the AI can clearly explain how it went from inputs to recommendations. 

Once we get past all those hurdles and the AI is actually treating patients, we’ll want to maintain oversite.  Is it keeping up with the latest research?  How many, and what kinds of, patients is it treating?  Most importantly, how are its patients faring?

I’m probably missing some that others more knowledgeable about medical education/training/ licensure might add, but these seem like a fair start.  I’d want my AI physician to excel on all those.

I just wish I was sure my human physicians did as well.

London cab drivers have famously had to take what has been termed the “most difficult test in the world” to get their license, but it’s one what anyone with GPS could probably now pass and that autonomous vehicles will soon be able to.  We’re treating prospective physicians like those would-be cab drivers, except they don’t do as well.

According to the Association of American Medical Colleges (AAMC), the four year medical school graduation rate is over 80%, and that attrition rate includes those who leave for reasons other than poor grades (e.g., lifestyle, financial burdens, etc.). So we have to assume that many medical schools students leave with Cs or even D’s in their coursework, which is performance we probably would not tolerate from an AI.

Similarly, the textbooks they use, the patients they see, the training they get, are fairly circumscribed. Training at Harvard Medical School is not the same as even, say, Johns Hopkins, much less the University of Florida College of Medicine.  Doing an internship or residency at Cook County Hospital will not see the same conditions or patients as at Penn Medicine Princeton Medical Center.  There are built-in limitations and biases in existing medical training that, again, we would not want with our AI training.

As for basing recommendations on medical evidence, it is estimated that currently as little as 10% of medical treatments are based on high quality evidence, and that it can take as long as 17 years for new clinical research to actually reach clinical practice. Neither would be considered acceptable for AI.  Nor do we usually ask human physicians to explain their “black box” reasoning.

What the discussion about training AI to be physicians reveals is not how hard it will be but, rather, how poorly we’ve done it with humans.

Human physicians do have ongoing oversight – in theory.  Yes, there are medical licensure boards in every state and, yes, there are ongoing continuing education requirements, but it takes a lot for the former to actually discipline poorly performing physicians and the requirements for the latter are well below what physicians would need to stay remotely current.  Plus, there are few reporting requirements on how many/what type of patients individual physicians see, much less on outcomes. It’s hard to imagine that we’ll expect so little with AI physicians.  

----------------

As I explained previously, for many decades taking an elevator without having a human “expert” operate it on your behalf was unthinkable, until technology made such operation as easy as pushing a button. We’ve needed physicians as our elevator operators in the byzantine healthcare system, but we should be looking to use AI to simplify health care for us.

For all intents and purposes, the medical profession is essentially a guild; as a fellow panelist on a recent podcast, medical societies seem more concerned about how to keep nurse practitioners (or physician assistants, or pharmacists) from encroaching on their turf than they are about how to prepare for AI physicians.  

Open up that guild!