Just Months After Jeopardy!, Watson Wows Doctors

kkleiner writes "Following its resounding victory on Jeopardy!, IBM's Watson has been working hard to learn as much about medicine as it can with a steady diet of medical textbooks and healthcare journals. In a recent demonstration to the Associated Press, Watson showed a promising ability to diagnose patients. The demonstration was a success, and it is the hope of IBM and many medical professionals that in the coming years Watson will lend doctors a helping hand as they perform their daily rounds."

Re:Idiocracy...

[Grygus]

You do realize that memorizing and regurgitating known information is the perfect skill for 99.9% of medical diagnosis? As long as Watson knows how to say, "I don't know" it will be as good or better than the vast majority of humans at this particular task.

Re:Idiocracy...

[similar_name]

However, they're still a ways to go before they can actually create new information.

This is true of most people.

Once they can do that though, that's when AI becomes a reality.

I always found it interesting that computers are never good enough until they can beat the best that humanity has to offer. Computers could beat most people at chess long before beating grand masters, but it wasn't until computers could beat the best human in the world that they were good enough. Likewise, Watson had to beat the best Jeopardy players before being good enough. So now, computers have to be better than the best doctor before being good enough. So, even if you make a computer that could graduate in the middle of a class of doctors, it won't be good enough until it can do better than them all. I just find it interesting as it says so much about us.

Wasn't this the promise of...

[Nutria]

LISP and Prolog-based expert systems 30 years ago?

Potentially Useful

[izomiac]

Personally, I seriously doubt that Watson will ever advance to being able to replace a doctor for non-trivial complaints. First of all, humans are better at image processing, so if a patient looks like death then they aren't going to ask questions to rule out minor complaints. Second, patients usually don't know how to describe their symptoms, and it's up to the doctor to make sense of what they're describing (keeping in mind that some exaggerate, some understate, and others outright lie). Third, clinical references are written for humans, so they often omit various "obvious" things (e.g. to get Lyme you have to have been bitten by a tick, which may not be very likely in Barrow, Alaska).

OTOH, I can see Watson being immensely useful on the back end. For example, which second-line blood pressure medications have been show to be highly effective with few side effects in 65 year old male caucasians who also have diabetes, and, of those, which has the best interaction profile with the other drugs this patient is taking? Clinical guidelines help, but they're obviously simplified and generalized. It'd take a human ages to research the literature to figure that out, but an AI like Watson could potentially do it in a few seconds. Such a tool could take a lot of the guesswork out of medicine.

Re:Wasn't this the promise of...

[brusk]

Not quite. The idea then was that we would teach the machines the rules, and they would follow them better/more cheaply than a human brain. The innovation here is that the system goes and looks at the published medical literature and figures out the rules on its own.

Better job than humans

[Oxford_Comma_Lover]

It will absolutely do a better job than a bad human. This should make a major difference in the long tail--i.e. things that aren't the obvious problem to the doctor, notably in second and third-rate hospitals. It will make procedural screw-ups a bigger cause of death and hospital problems as compared to medical malpractice. (I'm not sure what the ratio is now.)

It will also make humans more dumb and less thoughtful over time. That is, diagnostic skills will go down as diagnosis becomes done more and more by computer. The excellent doctors will still be excellent, but there will be even *less* requirement to really *think* about a problem than there is now.

Re:Wasn't this the promise of...

[ArwynH]

IIRC my AI classes correctly, those systems worked. At least they had a very high accuracy, higher than most doctors. The problem was not technical, but legal. Who do you sue if the computer gets it wrong?

Which makes me wonder: will this system will fair any better?

Re:Potentially Useful

[Dachannien]

First of all, humans are better at image processing

Medical image processing is a rapidly advancing field, but it's not a lack of technological advances that will stand in the way of automatic diagnoses based on x-ray/CT/MRI. Instead, the threat of malpractice will require that doctors manually inspect the images and render a diagnosis. Otherwise, you could probably expect automatic diagnosis from medical imaging in the next decade or so (or sooner, depending on what kinds of illness or injury you're looking for).

Malpractice is such a threat to medical imaging technology that a lot of medical imaging system manufacturers are afraid even to implement simple noise reduction techniques. If a cancerous spot gets removed as noise from an image and the doctor misses it, it could be a multimillion dollar suit against both the doctor and the manufacturer.

OTOH, I can see Watson being immensely useful on the back end.

Well, he did always want to be a proctologist when he grew up.