Microsoft claims their new medical tool is “four times more successful than human doctors at diagnosing complex ailments”. Unsurprisingly, what they meant by “diagnosing a disease” was the thinking-hard part, not the inputs part:
To test its capabilities, “MAI-DxO” was fed 304 studies from the New England Journal of Medicine (NEJM) that describe how some of the most complicated cases were solved by doctors.
This allowed researchers to test if the programme could figure out the correct diagnosis and relay its decision-making process, using a new technique called “chain of debate”, which makes AI reasoning models give a step-by-step account of how they solve problems.
If and when deployed, how likely is it that these algorithms will get a query comparable to a New England Journal of Medicine case study? Most doctors don’t reach those levels of perception and synthesis, let alone the general public.
I will have more to write about this soon (ha!), but until the stars align for an extended writing session here is a good opinion piece from FT’s John Thornhill about why LLMs may not be all that great for lay people dabbling in, for example, medicine:
When the test scenarios were entered directly into the AI models, the chatbots correctly identified the conditions in 94.9 per cent of cases. However, the participants did far worse: they provided incomplete information and the chatbots often misinterpreted their prompts, resulting in the success rate dropping to just 34.5 per cent. The technological capabilities of these models did not change but the human inputs did, leading to very different outputs.
The emphasis is mine, because it is a neat summarization of what I wrote 2 years ago. Humans are unique not because of what’s inside our heads but because of how we interact with the environment. There will be no artificial general intelligence until that problem is solved.
Once a decade, I am obligated to read a book from Eric Topol. Ten years ago it was during a rotation at Georgetown where they were handing around copies of The Creative Destruction of Medicine like candy. Of course, if those books had truly been candy they would have been of the sort that quickly congeals into an inedible hard lump because nothing in The Creative Destruction… aged well.
Well this year Topol has a book out on aging, and if it weren’t for some high-profile endorsments I would not be paying it two cents. But then I saw Nassim Taleb praising its rigor and scholarliness, highlighting as an example that Topol cites multiple trials for each claim. One can hope the trials he cites actually back up the claims, and to confirm that is indeed the case I now have Super Agers on the pile. Kindle version only: physical space in our library is too precious for Topol.
Noah Smith writes about health care costs:
So overall, health care is probably now more affordable for the average American than it was in 2000 — in fact, it’s now about as affordable as it was in the early 1980s. That doesn’t mean that every type of care is more affordable, of course. But the narrative that U.S. health costs just go up and up relentlessly hasn’t reflected reality for a while now.
And he has the data to back it up, though some of it feels like playing the denominator whack-a-mole. Interesting regardless. (ᔥTyler Cowen)
Yes, investigator-initiated clinical trials take time. But rather than back-patting and boasting about how it can still be done despite the setbacks, why not propose solutions for how to speed them up? I made a few off-the-cuff suggestions but you can also find serious efforts on that front.
A major entry in the Annals of Zombie Medicine must be screening for prostate cancer in men age 70 and above. Recent events had Nassim Taleb asking whether one could detect aggressive prostate cancer early, and one could, but… Indeed, this kind of screening has been singled out as something not to do for more than a decade, and yet:
Prostate screening in men ≥70 has not reached a 50% reduction in use since the 2012 guideline release.
Meanwhile, a full one-third of adult Americans is not doing the kinds of screening that are recommended, probably because they involve poop.
If all we had to do is trust the scientific method, why does homeopathy still exist (but not lobotomies)?
Another good podcast episode: neurosurgeon Theodore Schwartz talking to Tyler Cowen. Dr Schwartz is a bigger believer in science than yours truly:
COWEN: Do you think there are areas of science, though, where the institutions are so screwed up that you don’t actually trust the product of what is coming out, and there’s some systematic bias in the ideas being generated?
SCHWARTZ: I think, yes, there’s always going to be politics involved, and we always come to any problem from a unique single perspective, and institutions are going to have their biases. Yes, that is true, but in the long run, the scientific method will figure it out, and there will be one right answer. That institution — whatever their bias is — will be proven wrong in the long run. Now, those people might be dead and won’t be able to apologize at that point.
The problem, of course, is even when the scientific method does figure something out, people still keep doing things the old way, and no, generational change does not help. Witness homeopathy, kyphoplasty, vitamin C for colds, and — more relevant to Tyler’s question — the amyloid plaque hypothesis of Alzheimer’s disease. Abandoning lobotomies was an aberration, zombie medicine is the rule.
Where have all the healthcare YIMBYists gone?
Today in titles that trigger me: Where is the YIMBY movement for healthcare?.
YIMBY is too simplistic of a concept to be easily applied to healthcare. It relies on a single dimension — how easy is it to build housing — and any proposed policy can be easily placed on the NIMBY/YIMBY spectrum. And since everyone can agree on where a particular policy lies on that spectrum, creating alliances is easy.
No such luck for American healthcare, where there are many dimensions: accessibility of new procedures (clinical trial YIMBYism), accessibility of approved treatments (insurance YIMBYism), accessibility of healthcare providers (practitioner YIMBYism), accessibility of MDs in particular (doctor YIMBYism)… And even there it is not clear what the YIMBY-equivalent stance would be. Does clinical trial YIMBYism mean you want more trials, quicker trials, or just more drug approvals and doing away with trials entirely? If you are a doctor YIMBYist, do you want to increase the number of medical schools? Residency and fellowship slots? Enable more foreign medical graduates to enter practice? All of the above? But then how do you deal with practitioner YIMBYists, who want to do away with most doctors altogether and delegate most work to physician assistants, nurse practitioners and, at the end of the line, large language models?
There is not a single person in America who would say its healthcare system is working, and yet it is clear why there is no unified front on how to fix it.
To be clear, I quite like the ideas brought up in that leading article. The five sample issues it names — breakdown of the direct doctor/patient relationship, unclear fees for service, frequent insurance switching, no room for insurer creativity, too much money spent on end-of-life care — are spot on. If I had to pick one thing where I would want to be a YIMBYist, it is to remove any direct influence of the federal budget on healthcare. A large pot of money leads to hypertrophy of every other part of the system which down the line lead to many of the issues above. But is that really a YIMBY attitude?But alas the issues in question are too complex to be boiled down to a YIMBY-equivalent jingo, and to emphasize that point the article has an addendum linking to a 10,000-word report on the topic which at a glance seems to be raising the right points but I couldn’t really tell you since I have a day job that doesn’t leave much time for reading 10,000-word policy papers.
Today’s FT essay on the rise of the anti-vaccine movement was a miss. Instead of asking why so many people lost trust in institutions it goes straight to politics: 10 paragraphs on Germany’s AfD, no mention of whether some of the people’s concerns were valid. With that, the movement can only grow.
A one-two punch on clinical trials from Ruxandra Teslo and Willy Chertman today: first their on-point agenda for clinical trial abundance as a guest post in Slow Boring, then Ruxandra’s longer essay which has been so thoroughly research that even yours truly gets a name-check. As I noted elsewhere, every US institution has made one bade tradeoff after another in how it conducts clinical trials to the point that it’s impossible to conduct a RECOVERY trial equivalent over here. That needs to change.