4 ms·

The silly thing here is that it’s very easy to train models nowadays, so the medical AI space is very noisy. With the right data and the right application, thi

The silly thing here is that it’s very easy to train models nowadays, so the medical AI space is very noisy. With the right data and the right application, this technology can definitely be helpful. However, it is very hard to sort out signal from noise since any reasonably competent undergrad can throw a convnet at some images and get an AUC. Actually, I have even seen a high school science fair where they did a medical computer vision application. I think this speaks more to the accessibility and ease of use of the technology that is out there than it does to the inherent limitations of the tools we have. I am reasonably confident that there is a group out there that was more careful and has something that might have been useful, but it’s just impossible for clinicians to know who to trust. I feel that both uncritical, un-nuanced hype (Radiologists will be replaced in 5 years!) and uncritical, un-nuanced closed mindedness (AI will never be clinically useful) is harmful for tackling the biggest problem the field faces right now: trust. Who to trust. When to trust.

mattnewport 5y ago
> I am reasonably confident that there is a group out there that was more careful and has something that might have been useful What is that confidence based on? It seems like the researchers here made a reasonably thorough effort to find "something that might have been useful" and only turned up two that were even worth further investigation: "She and her colleagues have looked at 232 algorithms for diagnosing patients or predicting how sick those with the disease might get. They found that none of them were fit for clinical use. Just two have been singled out as being promising enough for future testing." So are you reasonably confident that these researchers just didn't look hard enough to find the good stuff?
- Grimm1 5y ago
  Why not? Plenty of researchers aren’t thorough. I’m not saying that’s the case here but it’s entirely plausible from what I know about academia and wouldn’t be a crazy claim to make.
  - Jweb_Guru 5y ago
    If a model did particularly well while almost every other model did badly, it would have gained attention in the community (not to mention, the researchers who worked on it would be actively marketing this!). Even if the researchers who worked on the paper weren't aware of it, it would almost certainly be brought up during peer review. The only way that such a model would not end up on such an extensive list is deliberate omission, not lack of rigor.
    - dontreact 5y ago
      I don’t think this is true. Medical practices change slowly and it takes time to build trust. It’s possible a covid model will become clinically useful, but it probably requires having enough buy in to run a clinical trial and then for the economics to line up so that using it leads to more money for hospitals somehow
- dontreact 5y ago
  I think the issue is not with the training step but with the validation step. The only way to build enough trust is to carefully run a prospective clinical trial, and before that you probably need enough retrospective evidence to get some sort of FDA approval. These processes take a lot of time and it’s not surprising that a year and a half after COVID-19 became widespread in the US, we still don’t have a fully validated model.
  - dontreact 5y ago
    In fact the claim the paper makes is that no model has been sufficiently clinically validated. Problem is: how do you get enough buy in from clinicians to run a clinical trial with YOUR model, when there are hundreds of crappy models out there? It takes a lot of time to run these studies for this reason to start with and then depending on the size of the clinical trial, it takes time for the trial to run it’s course.
iamstupidsimple 5y ago
I'm not sure if this has been attempted, so please do correct me, but I would very much like to see AI models designed to assist rather than give coarse classifications like "X disease detected". The best example I can think of would be creating bounding boxes around specific tells, or listing specific markers that a clinician would be looking for themselves, leaving the final decision up to them.
- dontreact 5y ago
  Nearly all FDA approvals for medical imaging diagnostic software are for programs that assist. It’s been done for decades.

mattnewport 5y ago

> I am reasonably confident that there is a group out there that was more careful and has something that might have been useful What is that confidence based on? It seems like the researchers here made a reasonably thorough effort to find "something that might have been useful" and only turned up two that were even worth further investigation: "She and her colleagues have looked at 232 algorithms for diagnosing patients or predicting how sick those with the disease might get. They found that none of them were fit for clinical use. Just two have been singled out as being promising enough for future testing." So are you reasonably confident that these researchers just didn't look hard enough to find the good stuff?

Grimm1 5y ago
Why not? Plenty of researchers aren’t thorough. I’m not saying that’s the case here but it’s entirely plausible from what I know about academia and wouldn’t be a crazy claim to make.
- Jweb_Guru 5y ago
  If a model did particularly well while almost every other model did badly, it would have gained attention in the community (not to mention, the researchers who worked on it would be actively marketing this!). Even if the researchers who worked on the paper weren't aware of it, it would almost certainly be brought up during peer review. The only way that such a model would not end up on such an extensive list is deliberate omission, not lack of rigor.
  - dontreact 5y ago
    I don’t think this is true. Medical practices change slowly and it takes time to build trust. It’s possible a covid model will become clinically useful, but it probably requires having enough buy in to run a clinical trial and then for the economics to line up so that using it leads to more money for hospitals somehow
dontreact 5y ago
I think the issue is not with the training step but with the validation step. The only way to build enough trust is to carefully run a prospective clinical trial, and before that you probably need enough retrospective evidence to get some sort of FDA approval. These processes take a lot of time and it’s not surprising that a year and a half after COVID-19 became widespread in the US, we still don’t have a fully validated model.
- dontreact 5y ago
  In fact the claim the paper makes is that no model has been sufficiently clinically validated. Problem is: how do you get enough buy in from clinicians to run a clinical trial with YOUR model, when there are hundreds of crappy models out there? It takes a lot of time to run these studies for this reason to start with and then depending on the size of the clinical trial, it takes time for the trial to run it’s course.

iamstupidsimple 5y ago

I'm not sure if this has been attempted, so please do correct me, but I would very much like to see AI models designed to assist rather than give coarse classifications like "X disease detected". The best example I can think of would be creating bounding boxes around specific tells, or listing specific markers that a clinician would be looking for themselves, leaving the final decision up to them.

dontreact 5y ago
Nearly all FDA approvals for medical imaging diagnostic software are for programs that assist. It’s been done for decades.