Software in development to detect / monitor infant pain
By Darren Murph 
posted Oct 19th 2006 6:35PM
posted Oct 19th 2006 6:35PM
Sure, we've got plenty of ways to inflict pain, but aside from obvious signs conveyed through body language, how can we be sure someone is actually hurting? In the case of infants, their facial expressions cannot be trusted to determine whether he / she is really aching, as hunger and desire for attention can yield very similar (and equally misleading) faces. Sheryl Brahnam, an information scientist at Missouri State University at Springfield, is currently developing software that has proven "90 percent accurate" thus far in truthfully differentiating between honest distress and false alarms. Brahnam's system, dubbed the Classification of Pain Expressions (COPE), analyzes facial signals such as "how scrunched up the eyes are, the angle of the mouth, and the furrow of the brow" to determine root causes. The system relies on a "neural-network learning algorithm" that has been trained on a database of 204 photographic images of 26 different infants taken during a "standard heel prick," which is widely known to aggrieve infants. Brahnam admits the software has "a ways to go" before ready for clinical use, but the ability to accurately detect pain could lead to quicker diagnostics in infantile issues, and probably keep clueless parents of whiny babies a tad more [Via MedGadget]
Good thing there is software in development. To me, all the babies in those pictures just look constipated.
Not too impressive. According to the linked article, only 60 of the images showed pain, meaning you get 70% by saying none of the images show pain. Further, it looks like the data used to fit the "neural network learning algorithm" was used to evaluate it, so the accuracy will be overstated. (This method of evaluation is akin taking a test, finding out which answers you got wrong and taking the test again. Naturally your score will go up, but it doesn't mean you've "learned" anything.)
I seriously doubt that the same pictures that were used for constructing the algorithm were used to test it... If so, I'd expect 100% accuracy. The language doesn't specify very well, but here's the key portion:
"Preliminary tests showed that the system was more than 90 percent accurate. This is remarkable, given how similar these expressions can look, says Brahnam. Even so, she is quick to point out the limitations of using such a small training set and still images instead of video."
If this is in any way being taken seriously, it would only be because they used the algorithm in new cases--a new set of real/non-real pain faces that would be fairly easy obtained. But note that this portion does NOT say that they used the same set to evaluate the algorithm. But maybe I missed something in another part of the article...
>>The system relies on a "neural-network learning algorithm" that has been trained on a database of 204 photographic images of 26 different infants taken during a "standard heel prick," which is widely known to aggrieve infants.
Seems to me the only 'prick' is whoever decided it would be a good idea to stab baby feet for research...
Farris, the heel prick is the way they get blood from newborn babies for doing blood tests; although the article doesn't explicitely say so, I'd assume they collected the facial photos during routine post-partum testing rather than inflict unnecessary trauma on infants.
Nice pun though. :)
Thanks, cc, I didn't know that they did that. Having never been in the delivery room (or in a hospital around that time), I was unaware that they drew blood from the baby. Thanks for informing me!
It probably depends on the local laws and health insurance, etc. but I think they usually take about 8 drops of blood and then send it to a lab to test for common diseases that are passed genetically... if you've ever seen the movie "Gattaca" it's actually a little eerie!
I sure as hell don't need a piece of software to tell when my daughter is in distress, and I would say most parents are the same.
Cranky, overtired nurses in hospitals pulling double shifts might not, so if this helps them, all the better.
you don't need a computer to tell you when a baby is pissed... the screaming says it all!
Yeah, it's pretty common to use only a portion of your known data to train the neural net, and then use the remaining portion to test it. Furthermore, the more similar the two different cases are, the better 90% sounds. There's often a 90% confidence requirement on neural network problems; if they could get better, there's often a more direct way of solving the problem. I didn't see them explain exactly how they used those 204 data, just that they trained on that set of data, so it really would make a difference whether they saved any cases purely for testing or not. We called it "cheating" if you only reported results that were in your actual training set. enough epochs (iterations of learning), and you can get almost any problem to give arbitrary precision. So I'd guess they didn't use the exact same data for the test results. But it's only my guess. :)
"I have soiled myself. How embarrassing."