AI equal with human experts in medical diagnosis based on images, study finds

[Michael]

AI equal with human experts in medical diagnosis, study finds

The team pooled the most promising results from within each of the 14 studies to reveal that deep learning systems correctly detected a disease state 87% of the time – compared with 86% for healthcare professionals – and correctly gave the all-clear 93% of the time, compared with 91% for human experts.

The next time I need a "second opinion", I might consider AI systems.

 

[public hermit]

I would have thought the difference would have been greater. 87% compared to 86%? You go humans! Still, who wants to take a chance on 1%?

 

[That Guy 11200]

The problem with these learning algorithms (they aren't true AI) is that they learn on specific sets of images from specific hardware or software. If they are put to work on images generated by different makes of the hardware or software, or potentially an updated version of the ones they learned from, then they can be hopelessly wrong.

Relying on these systems, when in many cases even their creators don't know what it's identifying in these images to make a diagnosis, could lead to a disaster. Especially if the expertise doctors currently have is lost because of over-reliance on them.

 

[Ophiolite]

Especially if the expertise doctors currently have is lost because of over-reliance on them.

That is a vital point, well made.

 

[Ophiolite]

The next time I need a "second opinion", I might consider AI systems.

The last time I asked my doctor for a second opinion she said I had the body of a ninety year old man. What I want to know is, how did she find out what I have in my freezer?

 

[The IbanezerScrooge]

The problem with these learning algorithms (they aren't true AI) is that they learn on specific sets of images from specific hardware or software. If they are put to work on images generated by different makes of the hardware or software, or potentially an updated version of the ones they learned from, then they can be hopelessly wrong.

Relying on these systems, when in many cases even their creators don't know what it's identifying in these images to make a diagnosis, could lead to a disaster. Especially if the expertise doctors currently have is lost because of over-reliance on them.

That and then there's always the Skynet scenario... 0_o

 

[Strathos]

That and then there's always the Skynet scenario... 0_o

Seems like a rather less effective method of exterminating humans than launching nukes, though.

AI: "You have a deadly disease, and the only way to cure it is to go... jump off a cliff. Yeah, really, so go do it."

 

[Speedwell]

But it's the future. Why do you think Dr. McCoy didn't seem to need to know very much about medicine nor Scotty about engines?

 

[timewerx]

AI equal with human experts in medical diagnosis, study finds

The next time I need a "second opinion", I might consider AI systems.

Thank goodness!!

I don't trust human doctors. I know a rather large percent of cases of misdiagnosis of major illnesses among persons I know personally!

The problem I see is due to absurd levels of specialization in the medical field.... And many doctors don't take their work seriously or only cares about making money... Too much specialization is causing too much guess work and signficant delays in diagnosis, the irony!

A computer AI may not need to be specialized in any certain field of medicine since it can store a much larger database of medical information and analytical data than a human mind.

 

[timewerx]

That and then there's always the Skynet scenario... 0_o

Why do people always think like:

AI >> Skynet, nuclear armaggeddon
Helping the poor >> socialism/communism

 

[FrumiousBandersnatch]

A computer AI may not need to be specialized in any certain field of medicine since it can store a much larger database of medical information and analytical data than a human mind.

At present, medical AIs are necessarily specialized, and they're trained on medical information from the very doctors you don't trust...

 

[timewerx]

At present, medical AIs are necessarily specialized, and they're trained on medical information from the very doctors you don't trust...

The human doctors providing data for AI, work under a much more relaxed environment.

They're not racing againt time to save a patient. And you collect information from probably a large number of volunteer doctors.

Human doctors in actual work environment are under much greater amount of stress, racing against time and really don't have time to cross examine results with a hundred more doctors. So if they don't have the right mindset for the job, the chances of making a mistake goes up a lot higher.

Big difference there. The only reason AI remains specialized is the lack of computing power with present technology. But we'll get there soon. Only a matter of time.

 

[FrumiousBandersnatch]

The human doctors providing data for AI, work under a much more relaxed environment.

They're not racing againt time to save a patient. And you collect information from probably a large number of volunteer doctors.

Human doctors in actual work environment are under much greater amount of stress, racing against time and really don't have time to cross examine results with a hundred more doctors. So if they don't have the right mindset for the job, the chances of making a mistake goes up a lot higher.

Big difference there. The only reason AI remains specialized is the lack of computing power with present technology. But we'll get there soon. Only a matter of time.

The AI training may be relaxed, but feeding data to the AI would be done by AI software and system specialists, not usually doctors themselves. The medical data that is collected to feed to AIs are typically the results of treatments given by doctors under working stress or racing against time to save a patient.

Regardless of computing power, generalising the capabilities of AIs is exceedingly difficult. There's no good reason to make generalised AIs for specialized tasks when they can be extremely effective in a restricted domain.

No doubt, in the long term, many specialized AIs will be connected together by higher-level administrative AI's to create compound systems (much like the functional architecture of biological brains).

 

[timewerx]

The medical data that is collected to feed to AIs are typically the results of treatments given by doctors under working stress or racing against time to save a patient.

Why would you try to feed the machine learning program with diagnostics data from doctors working under great pressure to save a patient??

Doesn't everyone have greater probability of making mistakes when working under pressure?

You'll be feeding it with inferior data quality even if you remove all cases of misdiagnosis.

You can instead collect historical patient medical test / lab results and hire multiple doctors to make independent diagnostics under a much more relaxed environment without time constraints and without having to save a patient.

You can even do both - collect diagnostics data and independent diagnostics to compare results.

 

[Ophiolite]

Doesn't everyone have greater probability of making mistakes when working under pressure?

I have not dealt with any medical situation, but in terms of business situations, personal difficulties or engineering problems I work more efficiently and effectively when under pressure. I know many others who are exactly the same.

Indeed, in the absence of pressure I sometimes have to manufacture pressure to in order generate my best work. In the absence of pressure I find it too easy to be sidetracked by incidentals. Pressure produces clarity of perception, accuracy of analysis and quality of decision.

 

[FrumiousBandersnatch]

You can instead collect historical patient medical test / lab results and hire multiple doctors to make independent diagnostics under a much more relaxed environment without time constraints and without having to save a patient.

I guess you could do that if there were enough doctors for hire that weren't working under pressure and stress because there aren't enough doctors...

 

[timewerx]

I have not dealt with any medical situation, but in terms of business situations, personal difficulties or engineering problems I work more efficiently and effectively when under pressure. I know many others who are exactly the same.

If you read my whole post, I also suggested to do both, i.e., collect actual diagnostics data and also collect patients' test / lab results for independent diagnostics analysis.

You'll have both data from actual cases and also data that is independently analysed in a controlled environment.

The beauty of this method is you can also see if cases of actual misdiagnosis can be rectified in a controlled environment and vice versa (if controlled tests can produce misdiagnosis against actual data).

It will also allow you to study the problem of misdiagnosis in greater depth and find out the contributing factors.

This way, you'll have a lot more data to analyse, for the sake of producing accurate medical diagnostics. Probably too much for a human to handle but not a problem for a machine.

 

[timewerx]

I guess you could do that if there were enough doctors for hire that weren't working under pressure and stress because there aren't enough doctors...

It's definitely going to be more expensive (probably a lot more expensive) gathering the best doctors around (and even from other countries) to be involved in a comprehensive medical diagnostics research for AI diagnostics.

You'll have to pay them, probably even more than they get from hospitals!

That much effort, yes... One of the biggest issues in quality of healthcare stems from medical diagnostics.

 

[FrumiousBandersnatch]

It's definitely going to be more expensive (probably a lot more expensive) gathering the best doctors around (and even from other countries) to be involved in a comprehensive medical diagnostics research for AI diagnostics.

You'll have to pay them, probably even more than they get from hospitals!

That much effort, yes...

OK; well good luck with that. As the old saying goes, "If wishes were horses, beggars would ride."

 

[Ophiolite]

If you read my whole post, I also suggested to do both, i.e., collect actual diagnostics data and also collect patients' test / lab results for independent diagnostics analysis..

That does not alter the point on which I was challenging you. In my personal experience and extensive observation of others, many people function best, with fewer errors, when they are under pressure. I strongly suspect that this is especially true of well trained, professional individuals.