Slashdot Mirror
Google DeepMind's AI Beats Doctors at Spotting Eye Disease in Scan (cnet.com)
DeepMind, Google's artificial intelligence business, is planning clinical trials of technology that can help diagnose eye disease by analyzing medical images after early tests showed its results were more accurate than human doctors. From a report: Published in the scientific journal Nature, the study claims that DeepMind, in partnership with Moorfields Eye Hospital in London, has trained its algorithms to detect over 50 sight-threatening conditions to the same accuracy as expert clinicians. It is also capable of correctly recommending the most appropriate course of action for patients and prioritise those in most urgent need of care. In a project that began two years ago, DeepMind trained its machine learning algorithms using thousands of historic and fully anonymized eye scans to identify diseases that could lead to sight loss. According to the study, they can now do so with 94 percent accuracy, and the hope is that they could eventually be used to transform how eye exams are conducted around the world. You might be wondering why we need AI to do this job that has up until now been carried out by medical staff. But diagnosing eye diseases from ocular scans is incredibly time-consuming for doctors due to their complexity. Due to the aging global population, eye disease is also becoming more prevalent not less, increasing the burden on healthcare systems.
Anything to knock these arrogant thugs in lab coats down a peg or twelve.
But it's great.
London so they have NHS unlike us where under the GOP system this can be used to quickly black list people.
If it's hard for a human to see whether the scan shows signs of a disease waiting to happen, what was the AI trained by? And by whom? Do we know that the eye scans are actually relevant to the diseases? This is the part that always strikes me odd in those "humans have a hard time to notice X, so we train an AI to do it" stories. If humans have a hard time telling whether something is or is not relevant to a certain disease, and if the AI can only be trained by humans because there is no other source of information available...
We used to have a Bill of Rights. Now, with the rights gone, all we have left is the bill.
Doctors are blind, but not because they can't see. They are blinded by their conflict of interest between patient care and controlling costs. This conflict of interest is never disclosed to patients, but is very real. A doctor will never tell you "hey I think you might have pre-cancerous cells, but I'm not going to order the test because it costs $2000 and it may very well not be pre-cancer. We'll just wait and see if an actual tumor develops," but 99 times out of 100, if a doctor doesn't order a test for something, this is exactly the reasoning.
My wife of 37 years died from cervical cancer that started as precancerous cells in her cervix that were detected but never acted upon by her physician. The doctor wanted to wait another 6 months before running any tests on it because, she said, the cells were not conclusively abnormal.
In the trial, the emails I had subpoenaed showed that the cells were in fact a high grade pre-cancerous legion, but an email from the insurance company instructed the doctor to put off further testing or a LEEP procedure because they were way over that quarters expenditure numbers and that it could probably wait 6 months without issue. I will never forget that email, "these things can always wait 6 more months."
Well, those six months cost her her life, cost me my wife, and cost our children their mother.
I am all for replacing doctors with machines and AI, so long as insurance companies are never allowed to tinker with the algorithms. Let's face it though, they will figure out a way.
Don't wait for a greedy insurance company to murder a family member to demand change. We need a medical system that puts health first.
By itself, the statement "94% accuracy" means nothing - without an understanding of the rate of false positives and false negatives in the diagnoses. Because of the generally low incidence of the specific diseases in the general population, better than 94% accuracy could be easily achieved by a black box that says "healthy" with respect to all diseases to be diagnosed. Of course, such a black box is totally useless, though it certainly could "transform medical care"!
Like many other forms of AI passing the Turing test, they are succeeding not because AI has improved so dramatically, though it has, but because humans have become such incompetent or lazy dolts.
In this case, did the physicians know they were competing? Physicians these days, at least in the US, spend like 7 microseconds looking at a scan. Errors are ubiquitous, not because the physicians are incompetent, but because they just can't be bothered doing a conscientious job.
Meanwhile, like most if not all neural net systems, there is no way to know WHY a particular diagnosis was made. A human can say, "Here, look at this right here. This is why we know you have age-related macular degeneration." A neural net says "Take my word for it, I trained in Philadelphia."
Comment removed based on user account deletion
we could probably automate 50% of the medical industry and still have a shortage of doctors.
That's a nice little made up statistic you have there. First off a lot of potential automation is refused by patients. They WANT a person to come in and talk to them about what they are experiencing and there is no way to automate this. Second, automation is only cheaper if you can do it in volume. Small medical practices don't have the money for expensive test equipment. There is a reason hospitals have the MRI machine and not your family doctor. This is neither good nor bad but just a reality of automation. Third, the cost overruns in the US medical industry have very little to do with automation and have a LOT to do with our ludicrous payment and management system. Fourth, the US medical system is actually relatively advanced when it comes to automation because there is money to be had by automating. If anything we have cost overruns because we have too much automation when we don't actually need it. Too much automation adds cost just like not enough does.
Every time we can move a diagnostic test from requiring 30 minutes of a doctor's time to 30 seconds of a computer's time, that is huge savings.
While it is true that any time savings is likely a cost savings, I think you may not appreciate how short a time doctors usually spend on a single case. My wife is a pathologist so I see some of this up close. I've seen pathologist go through 100 to as many as 300 cases in a single 8-12 hour shift. And for non-routine cases a computer likely wouldn't help shorten things very much. The actual diagnosis time is actually quite short - usually seconds to minutes. A lot of lab work is already highly automated. It's the gathering and compiling of all the data to do the tests and render diagnosis that usually takes the majority of the time and cost and labor. In my wife's job cases have to be accessioned, go through gross dissection, tissue preparation, histology and then finally looked at by the doctor. The doctor's percent of the time spent might be 5-10% of the total time spent. (I'm not even talking about the clinical time to take the tissue sample, transport it to the lab, and the paperwork and billing) Each of those steps has a person involved. The costs aren't the doctor's time but the legion of support staff needed to manage the equipment and paperwork.
These stories are often spun as computers taking over a doctor's job, when they really should be thought of as productivity enhancements.
You know who isn't worried about them? Doctors. You are quite right that they aren't and likely cannot be replacements for doctors. Much like the PC on your desk they are just tools to make them better at their job. And that's a good thing if we use them right.
Comment removed based on user account deletion
I think this could be a great tool for health care professionals. Basically the AI could suggest a series of possibilities and the Dr could use that as a verification or check. It's like they have a research assistant who can read through all the literature and suggest diagnostic or treatment or areas of interest. I wouldn't want to see this replace knowledgeable Doctors but assist them and make them more consistent, up to date and efficient.
Any Dr could be House with something like this. :)
Google beats doctor... but your doctor won't share your data with companies all over the web and start advertising to you based on what condition it discovers and trying to take advantage of you financially based on what ails you. (they will just milk you a bunch for your visit instead).
"That's the way to do it" - Punch
You can hide the automation from the patient.
Sometimes though not always. And you can't hide the fact that the doctor isn't there talking to them.
They can spend the time they would normally use to do so doing something else productive if the computer does it.
I guess I'm not making my point clear. The opportunity for time savings isn't generally in diagnosis. That's rather efficient in quite a lot of cases. It's on the administrative side of things that is where the real time burden is and where the opportunity for automation really stands out. Digital medical records, more efficient billing, reducing the need for office staff, etc. If we can improve treatment for reasonable costs by all means but we're not really saving doctor's time by making a diagnosis that used to take 2 minutes take 1. What would really help is to help them eliminate the time they spend on paperwork which is often several multiples of the time they spend actively taking care of patients (including diagnosis).
Where they find that after a few years of more usage that the system actually doesn't help at all.
A few days after this post, IBM shows low promise of success using AI for medical diagnostic. Why such contrast?