> Still, I'm thinking that as it improves, it's going to show that doctors are not that good at their job on average, and that's going to be fun to watch.
Medical AI is trained on labels generated by doctors. Can you explain how it will exceed the performance of doctors on average? Are you assuming that the labels will be generated by the "top x%" of doctors? If so, how will you identify those individuals? Or is there some other mechanism you're expecting to improve the performance?
I’m not sure, but it removes the conflict of interest (and subsequent gatekeeping) created by radiologists who don’t want technology to automate away some of their work.
There's a lot to unpack here! In the USA, the ABR regulates human radiologists via board certification. Medical technology is traditionally regulated by the 510K, PMA, and de novo pathways at the FDA. Of course, these products still have to demonstrate value in order for major stakeholders (hospitals, radiology practices) to purchase them. And using these products does not absolve the ordering doctor, the radiologist, or the hospital of legal liability for misdiagnosis. In fact, IANAL and this is a somewhat novel area of the law, but any AI product that functions as a drop-in replacement for a radiologist might be held liable for misdiagnosis that leads to harm. These liabilities could become quite large for a product deployed at scale (a single misdiagnosis causing death can lead to a settlement in excess of 10 million dollars). In summary, there's quite a bit more to the issue than simple "gatekeeping." It might be appealing to blame radiologists for these issues, but there's a much larger system at work that's designed to ensure quality and safety for patients. This is not AdTech - lives are at stake and people can get hurt. Now, this definitely comes with a cost to innovation, but it's going to take more than just a few MD's to reinvent the economics and law of computers practicing medicine on people.
I have very little experience with the existing medical establishment, so forgive me if I've imagined a scenario that is not relevant. I would like to be able to go to have a scan done, receive the scan data, then have the choice of submitting that to a radiologist of my choice or to an AI service of my choice that can read and recommend next steps. It seems like such a scenario is stifled by the existing way of doing things.
My experience with getting a sonogram was that the sonogram wasn't that expensive, but getting it read was hugely expensive. I understand that there are issues of liability, but it's really frustrating that I'm saddled with a high deductible healthcare plan where access to useful medical stuff is stuck behind 3-4 digit costs. Want antibiotics, inhalers, ADHD meds - all of which are pretty cheap in generic form? Pay $100 to the doctor for the privilege. People have very little agency in this system.
I guess at the end of the day I'd like to see open data (I'm able to get the images/data from all kinds of scans & diagnostics), and some kind of transparent system for submitting my data for diagnosis or analysis. There may be caveats & waivers, but I'd be willing to pay $10 to an AI service to tell me "You definitely need to consult with a radiologist based on the data presented" before I pay a radiologist orders of magnitude more to tell me that everything looks OK.
You're making some very interesting and valid points. You've correctly identified that when receiving bundled services you lose the ability to negotiate or comparison shop on the basis of price. This is unfairly combined with a legal presumption that when the healthcare system generates a bill, the bill is valid until proven otherwise. Now, although we probably need more physicians, I don't think the limited supply of physicians is the primary reason for this situation - it's more due to increasing market concentration of health insurers on one side and hospital systems on the other, leading to regional monopolies that don't compete on price. In fact, with the decline of private practices and the rise of hospital systems, physicians receive less than 10% of all healthcare revenue. I can see that you'd like to unbundle your healthcare and regain control over prices, and I think that's a very reasonable thing to want.
Interesting, can you give an example of a radiologist hindering progress? You make an interesting point about radiologists setting practice standards - what alternative do you propose? You may also want to consider that radiologists don't determine practice standards in a vacuum - they have to serve the needs and expectations of their clinical colleagues.
Oh that's a good point, I should have been more clear (and honestly had forgotten BrainWave, you can option that with their DTI fibre stuff).
These systems are still not on the scanner (unless some of the latest acquisition stations support it? Still not on most of the deployed stuff) so typically another workstation is needed at least, this package does happens to be from the vendor. For anyone reading who finds this confusing, you can license on the machine the ability to do the "pulse sequences" you need, but it only handles the raw data - you typically need another step to process that into images you can use.
Otoh most of the people I know doing this on GE scanners used something else, but that could easily be sampling bias.
Awake cortical mapping has much better resolution than fMRI, and it avoids image registration issues, but it requires a great deal of planning and patient motivation. If it turns out that the language center is on the opposite side from the area of the surgery, then awake mapping might not even be necessary. This is why fMRI is often used as a planning step before awake cortical mapping during the actual surgery.
The other thing to consider here is unintentional media bias - when a lay publication wants information they tend to go to tier 1 research & academic institutions; exactly the sort of place that has all the latest new toys. This is often very far from typical practice. So people outside of healthcare hear about all sort of things that are a decade or more from general usage, even if they get there eventually.
From the outside, it's probably difficult to understand how conservative medicine is generally. Even effective new practices often take decades to become really mainstream.
> [...] exactly the sort of place that has all the latest new toys. This is often very far from typical practice. So people outside of healthcare hear about all sort of things that are a decade or more from general usage, even if they get there eventually.
To give a bit more context to my anecdata.
The surgery I was talking about was performed 10 years ago. In Poland. In city with around 700k people. In probably the best hospital for such surgeries in this city with only one other hospital also doing neurosurgery in this city but specializing more in the spinal surgery. The neurosurgical ward back then was in dire need of renovations so not exactly shiny new place that has funds for the best toys.
As I said it looked pretty routine back then.
She did not have fMRI for her following surgeries because tumors were not close to the speech center (opposite hemisphere, then frontal lobes).
Interesting; I've much more exposure to US. Overall, I'd say there are sites that do it pretty routinely, but far more sites that basically never do it. And others that don't even have the equipment/licenses if they wanted to. For what it's worth, overall neurosurgeons do much more spine surgery than brain. In the particular case you mention, I've seen a lot of mapping and/or testing for surgeries near eloquent function, but little fMRI. Outside of that, basically unheard of.
I'm not on the outside of medicine. Significant practice variation exists, but complex neurosurgical conditions are usually managed at academic centers and fMRI is used for planning in neurooncology [1] and epilepsy surgery [2].
Fair enough, I didn't mean to suggest you were I was speaking more generally but worded that poorly.
My comments come from systems supporting thousands of clinical neurooncological procedures (i.e. tumor resections) in planning and execution with very little interest or utilization of fMRI beyond a handful proponents and their sites. Quite literally barely on the radar of most of the neurosurgeons apart from occasional papers, and some of them are quite negative about it also.
I could have an inaccurate picture of the breadth of clinical practice, and it's certainly a couple years out of date, but I would be very surprised to find a huge upsurge of usage outside research had happened.
It is certainly the case that the articles claim of "transformation" hasn't happened in that space.
If the article had instead claimed that some of the trickiest cases tend to have fMRI done (true, surgeons will take all they help they can get trying cases that otherwise might be inoperable) or that they are a feature of high profile academic sites (also mostly true) I wouldn't have objected.
I respect your experience, and I'll concede that fMRI for pre-surgical planning in neurooncology is used at only some centers. If you worked mainly in neurooncology you may have missed some of the uses of fMRI for epilepsy surgery. Thanks for the discussion.
> This is why fMRI is often used as a planning step before awake cortical mapping during the actual surgery.
That must have been the case.
I guess surgeons were trying to do whatever was in their power to not damage her speech center. Right after operation she had a few brief seizures when she lost ability to speak for a minute or two, but as the damage healed they quickly stopped.
Thank you for sharing your story. Although fMRI has low resolution, it can be very useful for surgical planning in certain clinical scenarios. For example, to determine hemispheric dominance: "Although the estimated percentages are of some debate, language is the purview of the left hemisphere in approximately 95% of right-handed people and 70% of left-handed people ... At MSKCC, language lateralization mapping is most often requested in right-handed patients with left hemispheric lesions, left-handed patients with left or right hemispheric lesions, or right-handed patients with right hemispheric lesions and signs or symptoms of aphasia." [1]
People doing related research are interested in it, and a few teaching hospitals etc., but that’s a really small fraction. it’s not really a factor in most clinical practice.
Thus “transformed medicine” is more than a bit of a stretch.
> it’s not really a factor in most clinical practice.
For a nice anecdata point - I’m an MR technologist and clinically I’ve seen it done once, I’ve done it once and I’ve heard a colleague mention doing one.
I’ve done them as research scans a hell of a lot of times.
fMRI is used to plan procedures that put eloquent cortex at risk. For example, tumor resections. Typically, fMRI is used for initial planning (whether the tumor is in the dominant hemisphere, how to approach the tumor). Then, in the actual surgery, function is confirmed using awake cortical mapping (e.g. stimulating different cortical areas with a bipolar electrode while the patient performs language tasks).
No, it is not. In my experience it is at best rarely used for planning, i.e. typically planning is done without fMRI. It will probably have DWI but maybe not tensor information (although that is growing).
Awake cortical mapping and/or task based references are very typical (near 1/2 of procedures, iirc). fMRI is not.
I'm not saying it's a bad idea or anything. I have a pretty good understanding of all the tradeoffs for fMRI but I'm not arguing for or against it. I'm reacting to the claim that is commonly deployed in clinical (not research) practice, which just doesn't match my (fairly extensive) experience.
So the articles claims of "transforming medicine" are aspirational.
source: Have worked directly in this market (clinical), hashed this out with many neurosurgeons,seen the procedures, etc. See also MR tech comment in this thread, in my experience very typical.
I don't really know what you are responding to here.
As I posited, there is research interest and a few people doing it clinically. Your links support this. What they don't support is the idea that this is typical clinical practice.
For those interested, here's a nice Tweet and video from a neurosurgeon at the Mayo clinic in Arizona demonstrating the use of both fMRI and intraoperative motor mapping: "Preoperative and Intraoperative mapping help with selection, safety and strategy."
"All Cell Press and Lancet journals are part of the UC agreement. For these top journals, UC’s shared funding model — where the libraries share the cost of open access publishing with authors — will be phased in, with all Cell Press and Lancet journals integrated no later than 2023, midway through the four-year agreement."
"a limited number of societies that partner with Elsevier for their publishing have chosen to exclude their journals from transformative agreements, so their journals are not eligible for either reading, publishing, or both under the agreement. A list of these exclusions will be available soon and linked from this page."
IANAL but HIPAA compliance comes from following certain policies and procedures (e.g. for encryption and account provisioning). These rules are necessary but not sufficient to guarantee security. As for cameras in hospitals, I have only seen these pointed at beds in specific scenarios (e.g. epilepsy monitoring) but obviously it is important to keep these video feeds secure.
Medical AI is trained on labels generated by doctors. Can you explain how it will exceed the performance of doctors on average? Are you assuming that the labels will be generated by the "top x%" of doctors? If so, how will you identify those individuals? Or is there some other mechanism you're expecting to improve the performance?