Your Computer Is Smarter Than Your Doctor

Screw Deep Blue, the computer on which you're reading this blog post can diagnose Alzheimer's--better than a doctor.

From a new study in the scientific journal Brain.

Experts taught a standard computer how to diagnose Alzheimer's from brain scans, and got a 96% success rate.

The accuracy of diagnosis from standard scans, blood tests and interviews carried out by a clinician is 85%.

reason pits man vs. machine here and here.

Comments

Jamie Kelly | February 25, 2008, 3:18pm | #

Wow, this is a fascinating and incredible revelation. It's more proof that ... wait, what were we talking about?
Jalapeno cheese dip is overrated. Bye, mom. And let's do something about these spiders.

Tacos mmm... | February 25, 2008, 3:25pm | #

I'm not really surprised that computers are proving to be good at brain radiology. The brain is really a very symmetric organ with relatively little non-pathologic variation from person to person and from day to day (exceptions being made for comparisons of brains at dramatically different life stages - baby v. middle age v. elderly). Computers should do a progressively better job at interpreting brain images in the near future. Other imaging sites will prove more difficult, particularly the abdomen and pelvis, as the gut is asymetrical with frequent variations in size and position.

Syd | February 25, 2008, 3:29pm | #

This isn't surprising. What's the rate of false positives?

Warty | February 25, 2008, 3:31pm | #

Machine learning algorithms are frequently better than people at separating high-dimensional data into distinct classes, so this isn't terribly shocking.

(Finally! A use for that Master's degree!)

K. | February 25, 2008, 3:34pm | #

For those who want to do their part by having their idle CPU cycles used for science (alzheimer research, protein prediction and design, etC), google:

Folding@home

Rosetta@home

Danke!

Marcvs | February 25, 2008, 4:13pm | #

Only smarter in the same way that your computer is smarter than you because it can do Fourier transforms more accurately than you can.

John | February 25, 2008, 4:20pm | #

This is the future of medicalcare, if we can keep the doctors guilds from killing it. We are finally hitting the point where are machines are getting so good that the productivity gains seen in other areas of the economy can actually happen in really advanced labor intensive areas like medicalcare.

old computer | February 25, 2008, 4:23pm | #

...uh...

Morat20 | February 25, 2008, 4:35pm | #

Machine learning algorithms are frequently better than people at separating high-dimensional data into distinct classes, so this isn't terribly shocking.

(Finally! A use for that Master's degree!)

What? You too? I'm midway through my thesis (no professor was doing anything I found remotely interesting with genetic programming, or I'd have just done extra coursework). Machine learning is insanely powerful, inside certain limits.

It's quite a chore to adapt it at times -- although perhaps it's just GP's that are so twitchy. (My thesis is on, effectively, data mining using GP's in a situation where the problem definition causes unwanted evolutionary pressures on the system. So, twitchy. )

Warty | February 25, 2008, 5:03pm | #

I'm mostly done with my thesis, which is looking at ways to improve brain-computer interfaces. Half of it is processing the EEG signals with wavelets, and half of it is looking at the performance of different classifiers on the spatial/temporal data. It looks like the wavelets are a big improvement over Fourier methods, but there's very little improvement in using nonlinear classifiers over a simple linear classifier. It seems that the data is too noisy to separate them with a complex surface, so the more complicated machines tend to overfit to the training set. So I know what you mean by twitchy.

How many dimensions is your data? I've never heard of anyone using GP for data mining. I'd try an SVM with several different kernels, but that's because I'm biased toward them after working with them for a year.

In any case, cheers to those of us who have delusions of smartness...that cause us to go deeply into dept and live in poverty for a few years. Stupid, stupid. *bangs head on table*

joshua corning | February 25, 2008, 5:19pm | #

Who watches the watchers of Alzheimer's diagnosers.

Seriously...how can we tell the computer is better then what humans do if humans can't do better then what they are doing?

R C Dean | February 25, 2008, 6:01pm | #

ways to improve brain-computer interfaces.

Faster, please.

Warty | February 25, 2008, 6:05pm | #

Turns out it's hard, RC. Really really really hard. So hard that I can't wait to get done and get a real job so I can afford to play with motorcycles and guns again.

Morat20 | February 25, 2008, 6:08pm | #

How many dimensions is your data? I've never heard of anyone using GP for data mining. I'd try an SVM with several different kernels, but that's because I'm biased toward them after working with them for a year.

Small -- I'm doing financial data mining. Cheap, complete data sets and problem definitions that are very easy to judge success and failure.

Or so I thought. It turns out that "simple" was more complex in practice. :)

Neu Mejican | February 25, 2008, 7:50pm | #

Seriously...how can we tell the computer is better then what humans do if humans can't do better then what they are doing?

Bootstrapping a gold-standard based on cross validation with an unrelated measure.

Oh, and getting a second opinion.

;^)

Neu Mejican | February 25, 2008, 7:58pm | #

Syd,

re: False positive rates

When group I and group II were combined in a single data-set, patients were correctly assigned to the appropriate group in 95.6% of trials with the leave-one out method using whole brain images (sensitivity 97.1%, specificity 94.1%).

Finally, when group I was used to train the data and group II was used to test, 96.4% of patients were correctly assigned to the appropriate group (sensitivity 100%, specificity 92.9%). Conversely, when group II was used to train and group I to test, 87.5% of patients were correctly assigned to the appropriate group (sensitivity 95.0%, specificity 80.0%).

Joshua Corning,
BTW, usually that second opinion comes from an autopsy.

Neu Mejican | February 25, 2008, 8:40pm | #

with relatively little non-pathologic variation from person to person

Relative to what?

Serious question.
I think these basic techniques are going to be very useful in future diagnosis of developmental disabilities.

I think the truth of this statement has a lot to do with what metric the computer is using to classify. There seems to be very wide variation in sulci/gyri patterns, and the implicit white matter connections associated with them. This would be important for classification of many impairments, but seems to have a wide variation in the normal population. Or am I wrong on this?