>Some radiologists think that AI will be really good for filtering out normal images, so that they only have to review anomalies. But I don't think that a model that detects diseases will be too successful, even if they manage to make it really work.

Radiologists say this until they get malpractice lawsuits. Seriously, I once injured my wrist and the first radiologist said it was just a sprain, but no surprise the hospital was using a outsourced...overseas provider for that quick analysis. The hospital had a inhouse radiologist review it as a course of business later on in a few days and the hospital panicked with phone calls to get me back in because there were numerous wrist bone fractures and they took the first step towards liability issues

Could an AI do it? Sure. But you sure as hell need a second opinion

Sure but why not take the best of both worlds? Make a ai analyze your wrist fracture and also a radiologist analyze it. If the results don’t agree then you need to re-analyze.

That would even drive down the possibility of malpractice.

A radiologist would have insurance in place for malpractice that indemnifies her/him. So your suggestion that a radiologist would be motivated to use AI to reduce this risk does not seem applicable to me. The idea to run AI alongside the human tool seems to be the current default suggestion these days, now that we have collectively realised that AI cannot do anything serious on its own. Pair it up with a human and the two can work together, with some handholding. But it is a hard sell, the software is now just duplicating a capability, and all of the issues that raises of potentially having two answers to one question. As a radiologist business, you would have to dedicate manpower to investigating when the AI has a different answer. Honestly, if I had a radiology practice, considering the points above, my answer would be no.

> modern machines can detect anomalies but doctors still learn how to interpret EKGs and double check what the machine says.

Uh, TBH no professional so much as glances through the automated EKG summary. It's utterly useless and could be deleted with zero consequences.

I've seen a lot of ECG results, and I always did my own interpretation, but I always found the machine to be accurate, which is unsurprising since distinguishing rhythms is very codified anyway. My only issue with the machine is that it's not really doing something difficult that needs doing. You still need someone that understands how to read ECG's and make decisions about patient management based on past and present conditions. That is the hard part.

I'm curious though. Can you elaborate on the types of issues you have encountered? What brand of equipment were you using?

It's actually improved a lot in the past few years with CNNs being used: https://stanfordmlgroup.github.io/projects/ecg/

No chance that a dinky ECG from 10 years ago is running a neural net inside.

If we can have cheap scanners or other instruments to go with the AI, I think it could easily appear as a screening tool that less qualified medical professionals (midwives, nurses, GPs) or even the patients themselves can use. Those people already do various rough screening tests and refer the patient up the chain of expertise according to the outcome.

> If we can have cheap scanners or other instruments to go with the AI

That's a big if already. Not to mention that a lot of times you don't want to give unnecessary radiation to patients, so they wouldn't be interested in making, say, X-rays more accesible.

> screening tool that less qualified medical professionals (midwives, nurses, GPs) or even the patients themselves can use

But how beneficial would that be? When you need an scan, it's probable that something weird is happening, and you'll probably need more tests than just a scan, and also some specialist that knows which tests to order and what conditions to consider.

Once you think of it in that way, it makes less sense. You'd be investing money, resources, training, time into a solution that might benefit only a small subset of patients (those with something weird but that can easily be discarded by a cheap scanner with high certainty without needing an specialist), without even being sure of what actual clinical benefits you would produce. It's far better to invest and research in other areas.

I was thinking more like ultrasound. If some crude scanner was cheap enough for every GP's office to have, and an AI could decipher the noisy signal which is perhaps too messy for a human to read, it could be used routinely like blood pressure measurement and stethoscopes are.

If it's a regular screening test rather than a response to the patient's complaint, then false negatives wouldn't be as much of a problem because at least it's better than nothing.

Ultrasound scanners are fairly cheap. But operating them is not easy, you still need training to use them and know what you're looking at.

> it's better than nothing.

Not necessarily, that's the issue with screening asymptomatic people. You have to balance the consequences and rates of false positives with the benefits of true positives. If ultrasound screening mostly catches indolent diseases, or those where catching them so early doesn't affect outcomes too much; and you have a lot of people going through unnecessary/risky tests, you might end up doing more damage with those screenings. It's not so easy.

The potential harm runs deeper than this even. Once a structural abnormality is found, the nocebo effect can come into play and essentially cause symptoms in asymptomatic people, which is a terrible result.

Also, one thing that is found does not preclude existence of other things to be found. The first discovered problem may attract so much attention that potential others are neglected.

Tl;dr: the first diagnosis tends to stick with the patient.

I know that can be an issue but like you say, there's a balance and I don't think we know where that balance is for the non-existent AI technology with unknown error rates.

Newborn babies undergo a whole lot of simple and inaccurate screenings. I guess those are cases where the harm due to inappropriate treatments is lower than that due to leaving diseases undetected.

Either way, I'd rather the technology exist and then people work out how best to use it, rather than the decision being forced on us by it not existing.

I had an abdominal ultrasound recently - the actual process of placing the probe and annotating the images seems like a pretty specific technical skill. I don't think any old person (or even a doctor) off the street can just rub the wand on you and find your kidneys for instance.

Indeed. US is a highly technical, specific skill. Even within a specific use case -- e.g. echocardiography -- there's a specific series of qualifications that cardiologists need to get in order to be deemed competent. The rotation of your hand in a specific position in transthoracic echocardiography can change the resulting ejection fraction of the heart -- the main functional outcome. It's a simple thing to see done; quite another for you to do yourself and make safe decisions.

That's the idea of using AI. So the operator doesn't have to be skilled at identifying things.

I have had an abnormal ultrasound and then a CT negative. If AI can cut down the CTs it's a win.

Your argument would be better if you stated that most diseases are diagnosed at an advanced stage when it's obvious. The rest are incidental and then we sometimes occasionally we get lucky. We don't routinely CT scan people because it's harmful to do so.

No disagreement here from my part, but I wouldn't be surprised if imaging artifacts could be detected/fixed by ML. E.g., there has been quite a bit of research in MRI boundary effects, although I wouldn't know if that turned out any practical solutions.

The issue with that is that misclassifying something in an image can have serious consequences. So unless ML is 100% accurate, you'll always want an alternative way to check if it's really an artifact or something serious, so, why bother making an accurate model when you'll still double check the results?

I don't follow. Why would anything need to be 100% accurate? Humans aren't 100% accurate, and there's no reason ML would need to be 100% accurate either.

The is a gigantic difference between a doctor looking at something in a scan and deciding it's an image artifact when it's not, subsequently missing a diagnosis, and an ML algorithm looking at that same image, deciding something is an artifact and removing it from the image. At that point you're not looking at an actual scan, you're looking at something different.

It would be interesting then to give the human a side by side comparison with the artifacts removed (and maybe shown in color), so that the doctor can more easily see whether he agrees with that.

The difference is that a doctor can try to explain his findings and collaborate with others. This includes not just looking at the scan data but also reading other test results and patient history. ML is not capable of this as far as I know.

>This includes not just looking at the scan data but also reading other test results and patient history. ML is not capable of this as far as I know.

ML is capable of doing that if you train it to do that.

Correct, the target should be instead better than Xth percentile humans across the majority of cases, with the failure modes categorized.

I don't get this perfection requirement that gets put on so many systems here. No system is perfect, the question is what failure rate you accept.

You're looking at this in isolation instead of in the context of a diagnostic process. Thinking of medical ML models in terms of benchmarks does not make much sense, you're interested in whether they actually change the process.

So yeah, maybe you get a model with 5% false negatives where humans have 10% false negatives. That's good. However, when you go to apply that model in reality, what happens to those 5% where model and doctor disagrees? First, we should think about the kind of disagreement. In most of those cases I'd bet that the doctors see something wrong but can't say what, not that the model says "this is bad" and the doctor "it's not".

Second, we need to think about what happens when doctor and model disagree. As the model is not 100% accurate, the doctor doesn't know if the model is mistaken or if they're the ones mistaken, so they'll probably order more tests anyways. If it can be something serious, it's worth it to do an extra test to make sure. They'd probably ordered those tests anyways if they weren't sure of what was happening, model or not.

So what did the model for a single test change? Did it really change the diagnostic outcomes? What's the actual benefit of the model? How much it's worth to get from 10% to 5% false negatives with a model for a single test if just adding more tests (say, with the same 10% false negatives) to the mix can give you a 1%, 0.1% false negative rate?

That's my point. Unless accuracy is really high, ML models are not going to remove uncertainty in diagnostics, few diagnostics consist of just a single test. Benchmarking models against human performance in a single test is not a metric that can drive implementation in the real world.

I agree, that it would not need to be 100% accurate. However, automation with low accuracy would induce other kinds of errors (e.g. complacency or loss of skill in doctors). Therefore, automation shouldbe significantly better than humans to justify their use.

We were talking not about about clear signs but about small signs that could be artifacts or actual issues. In that case you are going to have to do some complimentary or repeated tests to verify, so the ML output saying something is an artifact will have no effect on those procedures unless it's 100% accurate.

But in general, these are matters of life and death, literally. You're still going to have someone looking at the images and verifying the output. That limits a lot the potential cost benefits of these applications.

I still don’t understand why decisions about life or death have to involve 100% certainty or 100% correctness. I would think that it’s the opposite, that when the stakes are high, you want anything that improves your odds. If your life were on the line, you’d accept treatments or diagnoses with much less than 100% certainty.

See my comment below for the same discussion: https://news.ycombinator.com/item?id=27433368

My point isn't that you need 100% accuracy. It's that a diagnosis is a process, not a single test. If your model applies to a single step of the process, and if it doesn't remove enough uncertainty about that step, you're still going to continue the diagnostic process and the model is not going to change anything really.

You said ML will have “no effect on those procedures unless it's 100% accurate.” I disagree with that, and think it’s a bit absurd—the 100% figure is not possible anyways, so you might as well stick some other impossible predicate in there and the statement is equally insightful.

E.g., “it will have no effect on outcomes until pigs fly.”

There will always be uncertainty, and uncertainty isn’t the only relevant parameter.

> There will always be uncertainty, and uncertainty isn’t the only relevant parameter.

But it's a really important one and a lot of medical ML research doesn't seem to address in the proper sense. A very simple example: a ML model that classifies lung nodules as benign or malign, with 95% accuracy vs 70% accuracy of regular radiologists. Very good, right? But for actual, real world results, you need to see how the patient outcomes change. If the patients where the model and doctors disagree were going to have extra tests or followup regardless of what the model says, the model is not actually offering anything new despite the increase in accuracy.

So no, it's not absurd to say that models that work for a single type of test need to be very, very accurate to actually bring changes to the procedures that are worth the investment. What matters is whether they actually change patient outcomes, and sometimes it seems like the ML researchers barely consider it.

Where I live it's not uncommon for doctors to consult the images with their colleagues whenever they come across something unusual, even if they are "pretty sure" of their diagnosis, because the consequences of an error can be severe - we're talking literally about life and death here. So the ML system would have to be as precise as 2-3 humans together in order to be considered "good enough".

The place where I feel ML can help is finding out new disease markers or help upscale the quality of a scan (something I believe Facebook is looking at)