With 'digital twins,' the doctor will see you now | Hacker News Reader

With 'digital twins,' the doctor will see you now

quantamagazine.org

39 pointsrbanffy22 comments

What purpose does this serve?

It's not clear to me that any biological system is understood at this level of detail. Any simulation is likely to be garbage-in, garbage-out. And even if it were fully understood, there's not enough computing power on the planet to model every atom and every molecule.

The map is not the territory, people. All models are wrong, but some are still useful. The ability to make simplifying assumptions is the whole point.

    "There’s a quantity called “wall shear stress,” which is just a frictional force on the wall. We’ve shown that when doctors can visualize that wall shear stress at different parts of the artery, they may actually change the length of the stent that they choose. It really informs their decisions."

This sounds pretty good to me? The point of fluid dynamics and any other science of physical process is about refining a set of those simplifying assumptions. You don't model every atom or molecule, that's what fluid dynamics is for.

This is just one of the thousands of cases of a journalists fawning over any ML/AI thing with no understanding of the garbage-in garbage-out problem, and all the other issues with ML. You are completely right, attempting to simulate an individuals coronary systems is 100% out of our wheelhouse right now. At very best it's going to tell us things we already know, that people with clogged arteries are at risk of a heart attack. At worst it's a grift on par with fully body scans on healthy individuals.

I'm not saying ML is useless, far from it, but this is reminding me a lot of Eric Yuan (the CEO of Zoom) talking about AI clones in meetings: https://www.youtube.com/watch?v=dKmAg4S2KeE

The only similarity I see between this and the dumb Zoom thing is that the term "digital twin" sounds like it should mean "that stupid idea where you send your clone to a meeting".

That's not what "digital twin" means though: it's been used to describe digital simulations of systems (like the circulatory system described in this story) for a long time: https://en.m.wikipedia.org/wiki/Digital_twin

I've seen the dumb AI clone idea called "personal digital twins". I don't like that it steals and dumbs down the existing useful term but we may be stuck with it.

A significant amount of health issues result from medical noncompliance. A digital twin might be useful for finding ways of manipulating patients into taking care of their health.

It makes consultants very very rich.

Digital Twins are part of every buzzword infested "Digital Transformation" effort in my industry.

I don't trust this buzzword, because I've seen it recently my industry (space). An industry where there are very good thermal, mechanical, kinematic, electrical(I could go on) models, all correlated to real-life testing, and usually an entire spacecraft simulator is available and used for operator training. There is complete traceability to every critical component. And even in this industry, in which a product has many correlated models made by different engineering disciplines and through a paperwork trail can identify the root cause of almost any issue, and then have the issue reproduced in the relevant model or a simulator "digital twin" doesn't mean "find a way to take all the available model outputs, test data, etc and predict a a failure mode". People have been working on doing this for decades. It ignores all of these available inputs and says "use AI and stuff to make a digital twin".

I think I may agree? What I see is Digital Twins with "AI and stuff" being promised as the savior of design onto manufacture then deployment and eventually operation. Draw block diagrams and software magically checks and flags errors which can then be fixed. Then, the model becomes "the truth" even when reality intervenes. Pretty tall order especially if the thing is the first or only one of its kind.

There's also an implicit bias towards everything being decomposed as if it were digital and cleanly separable. Reality is much more analog and interconnected.

It is worth noticing that, while the article might be trivializing the topic a bit, the researcher won an ACM Prize in Computing so I would wager she knows what she is doing. Digital twins may be a buzzword but there is also very very solid research into them, pretty much like AI/ML.

Twin Health uses the same concept for improving metabolic health

https://usa.twinhealth.com/

I was initially thinking that teeth would be a good one for dentists to be able to check using a digital twin.

A patient simply walks in, have their teeth scanned, and a dentist can evaluate problems quickly from their desk.

That pretty much exists right now. My dentist did a 3D scan of my teeth (with a fancy wand that went in my mouth) and showed me a rotating model on a screen while talking about what they wanted to do.

That's like 40% of all dental visits. Take x-rays and conventional photos of teeth, put on computer.

Not related,but some of the ugliest json i ever saw behind a rest API was produced following the digital twin approach. Those buzzwords are a signifier of cluelessness and clueless birthw horrors.

This is kind of cool. Like weather prediction.

I wonder if we can do things like take a digital copy, then try:

- eat sugar for a week

- run a marathon

- eat "healthy"

- act "normal" (control)

Then do some comparisons. Might be enlightening.

I'm curious to know how far into the future we'd have to simulate to get meaningful data. Obviously it would matter how extreme the case is, using your example of eating sugar (I assume you meant only sugar) for a week, might quickly show us consequences but then how long would eating "healthy" need to run for?

Will that be the limiting factor in the future? I can hear it now, "I'm sorry, your insurance only allows for a 3 day simulation."

Ideally, all the way. Patients would probably care quite a bit about the bottom line number of overall life expectancy impacts of various choices.

Now, that's entirely out of our scientific and computational abilities no matter what anyone says. It likely will be for a few decades, at minimum. Though by that point who knows what kinds of new medical technology will exist. I wouldn't be surprised if this question is irrelevant by the time we can answer it. Just idle speculation though

>Ideally, all the way.

If it turned out that your body had some kind of problem that would guarantee death within some margin of error around a known date, and there was no cure or solution available yet, would you want to know?

Curiosity wise I think I would with the reasoning that it'd allow me to budget the time I have left more wisely and spend it on better, more satisfying things. But then thinking about it more, I'm not sure if I'd really want to. I think the stress of knowing exactly when I'd die would probably end up significantly impacting my life and causing health issues that would move up the death date ahead of when it really should be. And I think I'd end up enjoying those "meaningful" moments less because they'd feel like I'm rushing through a checklist of things to get ticked off before dying.

If it is possible to know I would want to know because if it is possible to know I would end up obsessing over it anyway. I wouldn't be able to forget that it is possible to know and my curiousity couldn't handle it.

I also don't think I am currently particularly enjoying my life to an extent where checklist like idea of rushing through a bucket list would make it worse.

Digital twins are very limited right now so they won't be able to simulate such actions with any useful fidelity. And what does it mean to eat "healthy" anyway?

If you want to know the effects of changing your diet or running a marathon then just go do it. Why would you need a simulation for something so simple and inexpensive?

Just wait until the health and life insurance providers take hold of the data and they start making decisions based on those simulations.

Do they not effectively already do this? More data provides more ability to measure risk.

Reply on news.ycombinator.com