Artificial intelligence versus clinicians: systematic review | Hacker News Reader

Artificial intelligence versus clinicians: systematic review

bmj.com

87 pointsdoener28 comments

This sounds about right to me intuitively. I've read a ton of ML/AI papers in the area, and the vast majority don't get past the "maybe this is an interesting idea, hard to know until it's tested properly". Easy to recall interesting ones, hard to remember good one off the top of my head.

This is fine; of course the standards for "is this publishable" are vastly lower than "is this clinically applicable", but it unfortunate how often the lay science press fails to make this distinction.

The other thing is that raw performance numbers, even if properly validated, are only a small part of the story. We've had ML techniques that did better nominal sense/spec than average clinicians for very specific tasks for 20+ years now, but the workflow impact, liability, etc. combined has sometimes limited the implementation, and even if you do roll it out the actual impact can be minimal.

Have a look at the Stanford Machine Learning Group projects [1]. What most articles miss in my opinion is that AI progress heavily depends in the data being available and that‘s where there are huge opportunities in the future.

[1] https://stanfordmlgroup.github.io/

They have a good group and are thinking/heading in the right direction, mostly not there yet I think.

I certainly don't disagree that there is potential for impactful use of AI in some areas of healthcare. Some of them just aren't going to practically happen without a) fundamentally reworking our approaches to data access and sharing and more importantly b) spending a lot of resources on quality labeling.

Regardless of whether or not those preconditions are actually met, we'll get a lot of breathless claims on small retrospective sets, of course :)

Medicine is one area where measurements are key in determining health outcomes. I don’t see any reason why an inference engine wouldn’t be at least as good as a human doctor. Although, self-supervised learning will need to become a reality, before these systems have any practical use IMO.

Don't misunderstand me - I think there is real potential in to make an impact in the area but there are systemic and cost issues to doing it well.

That said, the problems are harder than appears to your typical ML grad student, and structurally we aren't there with data. There is no reason currently to think self supervised with "get there" in the near term, but even if you've assumed infinite labeling resources the data sets are for the most part still too disjoint and unavailable. Some strides have been made in some areas recently, but a long way to go. The algorithms part isn't necessarily that interesting really, building up the supporting infrastructure is the real problem.

Very interesting paper, thanks. It was time someone looked into it. I can't take these equally populist and unsupported reports anymore that some neural network is doing better than a medical specialist with decades of experience. It may well be that this applies in individual cases. But obviously there is a lack of proper studies and unbiased evidence.

> It may well be that this applies in individual cases.

Well from a machine learning perspective it will probably be in all cases where enough, qualitative data is available and not a lot of contextual knowledge is necessary that cannot be represented in the data. My guess is that this will be the case in a vast majority of cases and it's just a matter of time - but I'm definitely biased as an ML researcher :).

It's still a "proof by induction". There is no reason to conclude from the individual cases that it works in all cases.

We also looked at this for Dermatology with similar conclusions https://onlinelibrary.wiley.com/doi/full/10.1111/bjd.18880

Very interesting, thanks. I also find it remarkable that most affiliates of both papers are from the UK.

That's the real value (going forward) of a national health insurance system treating most people and having a financial interest in keeping them healthy long-term.

A few thoughts below as a medical student with some ML background. I'm quite interested in hearing what others think: 1. The majority of clinical ML applications are still focused on narrow (e.g. binary) image recognition tasks. This makes sense because it's where CNN's excel. Also, many of these tasks are well reimbursed. I expect the performance to continue improving as datasets and annotations improve. 2. There is less activity on broad image recognition tasks, like reading a CT Abdomen and Pelvis. Here the domain of possible diagnoses is rather large. For example, patients sometimes have unusual anatomy from prior surgeries. Or, a cancer may obliterate common anatomical landmarks. An experienced radiologist or surgeon can figure out this anatomy in a few minutes, whereas ML may need new approaches to get to this level. 3. The history and physical exam are clinician-generated, not just for dataset annotations but for every new patient. This is notable because the history and physical exam determine the majority of what a patient needs. It is not only a matter of knowing which questions to ask, it also requires follow-up questions and a degree of mental processing before findings are written down in a note. A lot of patients will not fully cooperate without patience and empathy. So "AI" replacing doctors actually trend perilously close to AGI. This is both interesting, and difficult. Of course, NLP of human-generated notes / unstructured data is a very interesting area, but it still requires a human in the loop.

As someone with your exact background, I agree with you on almost all points. However I think that the common notion is less that AI will/should replace physicians but rather that AI could help physician to free time that they could then spend with patients. So the ongoing human involvement is an almost explicit goal (outside of radiology and pathology at least).

On another note and speaking from personal experience, NLP as well as patient information and note processing is something that might be less glorious than AI analysis of MRIs for example but which would have a far greater impact on physicians’ day to day life. Most residents spend like 50% of their time fighting applications with some of the worst UX I have ever seen just to get some basic information, log something or schedule a simple procedure.

Another medical student here, also with a very similar ML/programming background. Like the other commenter, I couldn't agree more with your analysis of the main issues plaguing current DL applications in medicine. When I was working on deep learning projects as a undergrad before coming to medical school, I naively assumed that solving a simple image classification problem for cancer detection would be enough. However, as one learns in medical school, imaging is only one component of an entire clinical vignette. Even before a patient undergoes a specific test, the history and physical exam really drive the initial protocols. Sometimes, without having this background knowledge, the classification from a simple program has no utility. A radiologist or pathologist typically has access to this information and can interpret this in the context of the clinical suspicion being put forth. I still believe AI has a role to play in easing this workflow, but "replacing" physicians will take a lot more than detecting some disease "better" or more "accurately" than a physician.

This paper references two tools for assessing the risk of bias: the Cochrane risk of bias tool and PROBAST. Does anyone have any experience with these tools? Are they something that are mainly only practical/useful in the medical field, or are they useful in other domains applying AI models?

The most common model that we are putting in front of FDA several times per week for many companies is the AUROC model with minimum FDA requirements for specificity and sensitivity. I have some concern that the commentators on this article aren't actually aware of how many of these ML based software packages are actually being cleared by FDA at the moment through the DeNovo and 510k processes.

There are a lot more systems going through 510(k) for sure, partially due to new developments, but partially driven really by marketing.

On the other hand there have been ML/AI systems since the mid-late 90s at least, and they haven't for the most part shifted clinical practice much. The reasons for this have not shifted radically in the last decade, although some are moving a bit now.

FDA approval to market doesn't mean anything beyond you have an argument that you haven't introduced new risks (safe) and in at least some cases you can improve something without making others worse (efficacy) but this may in a very narrow indication.

I guess my point is, the purpose of the FDA has never been to ask "does this really work better? should it be the new standard of care?", but rather to balance the risk in new technology or drugs against the potential reward. Even with a PMA it's risk reduction, not proof of how it will play out in the wider world if interactions and practice.

Haven't had the time to read the paper but spoke a lot about something similar with a friend.

In terms of crisis, the medical system just doesn't scale. You think leaders were worried about running out of ventialtors?

No! You can start pumping them in hundreds of thousands if you really need it.

The problem is that you have only so many doctors and trained nurses to work with. Limited capacity. Training takes 10 years.

No scalability.

It doesn't matter where the tech is right now. The world will push 100% for whatever can be done via apps, sensors and unskilled workers.

Check out Jonathan Rothbergs vision for a diagnostic kit by everyones toothbrush

The sad truth is that most patients cannot communicate their symptoms clearly and a computer will fail them. More importantly, the nonverbal communication, the compassion, and the reassurance from personal interaction are what drives the demand for personal service. Just look at the number of negative reviews on health grades. Many people are so unhappy with the way they were treated, even though the actual care they receive in this country is good- access to CT scans, antibiotics, home nursing, hospital service, etc..

If you make a comparison with the same amount of time for both, that's true. But software can ask questions for an hour. A doctor won't.

Important study to do, but feel like I have to assert that in principle, ML is not as useful for diagnosis as it is for aggregating longitudnal data because the consequences of a false positive/negative in the individual diagnosis are life threatening, and even if it catches something doctors don't, it just means a doctor has to look at them.

However, across a population, nobody is being harmed by the ML analysis, and it can pick out patterns a person could not at speed and at scale. Spending research to get your individual diagnostic ROC curve a few percent higher is a waste of time if the consequences of it being wrong are greater than the time savings it provides.

It's fine for veterinary medicine, and maybe prisoners and keeping people with hidden symptoms out of crowded places, but for regular people, using ML has got a host of ethical problems that aren't being dealt with at the policy level. You can use rules engines (literally, DROOLS) instead of non-deterministic ML schemes to do diagnoses and prescriptions.

Have said before, ML is only useful for problems with asymmetric upside, and it is worse than bad for ones with asymmetric downside.

> It's fine for veterinary medicine, and maybe prisoners and keeping people with hidden symptoms out of crowded places, but for regular people,

Excuse me but did you just compare the application on prisoners and animals versus whatever "regular people" are supposed to be? What the hell?

Countries treat prisoners as less than human everywhere in the world. It's awful, but it's real, and substandard health care from machine learning applications is precisely the kind of thing states everywhere would likely use. Most people are not so easily shocked by this.

Reply on news.ycombinator.com