Is there valid physics behind the bodybugg? The bodybugg is a faddish gadget whose marketers claim it can measure your body's daily energy expenditure. Their sales literature says:

As reported in the British Journal of Sports Medicine, 2008 edition, the bodybugg® armband achieved average accuracy rates of 90% in studies comparing calorie expenditure measurements from the bodybugg device to calorie expenditure measurements in a metabolic chamber (the "gold standard" of calorie expenditure measurement).

"Accuracy rates of 90%" is obviously not how anyone with any level of science literacy would describe error bars. Their FAQ says:

In addition to measuring motion and steps, the bodybugg is able to see how much you are sweating, your skin temperature and the rate at which heat is being dissipated from your body.

They also say it doesn't measure your heart rate.
Is there valid physics behind this device, and if so, what principles would it be based on?
Measuring "motion and steps" would seem to involve an accelerometer, but it seems unlikely that that would be able to tell the difference between, say, walking on a horizontal sidewalk and walking uphill.
How can they separately measure skin temp and rate of heat dissipation? Maybe there's a thermometer in contact with your skin and one sensing the air temperature?
It's not obvious to me how one could combine the mechanical data with the thermodynamic data.
 A: The short answer is yes.
Perhaps at the the time this question was asked the website was formatted differently, but they have a short paper I found almost immediately:
http://www.bodybugg.com/pdf/wp_accuracy_ee.pdf
Being skeptical of anything that's not in explicitly in a recognizable journal, the first thing I did was look into the author list.
The first author, David Andre, seems to be reasonably credentialed, see: http://davidandre.com/cv.pdf. 
The company he was director of research for, BodyMedia, presents at IEEE and has published research through them as well IAAI (see: http://www.aaai.org/ocs/index.php/IAAI/IAAI-11/paper/view/3533/4018 and Wearable sensor badge and sensor jacket for context awareness).
A general bibliography for BodyMedia is available here:
http://www.bodymedia.com/Professionals/Bibliography?whence=
While I can't claim that the numbers they use in marketing are accurate, it appears that there is a growing body of respectable research into physical activity monitoring and they're doing their part.
It might be worthwhile to look over some of the naive technical specifications if they weren't available to you before:
http://www.bodybugg.com/pdf/armband_specifications.pdf
Hope this added to what you had already found, assuming you still care to know!
[Edit by bcrowell.] The following is my summary of some of the information from the paper on the bodubugg web site. It measures heat flux, galvanic skin response (sweating), and acceleration. Unfortunately the science behind the thing is secret:

The algorithms are created using a proprietary algorithm development process
  that utilizes a data driven machinelearning approach.

So third parties who have tested it have only been able to test it as a black box. There is a little general information about how the algorithm works and what it can and can't do:

Essentially, the algorithms break down a person’s activity into fundamental activities of walking, running, resting, sleeping, resistance activity such as weight-lifting, lower-leg motion such as stationary biking, motion caused by external forces such as driving a car, exercise combined with external motion (road biking), etc...
  The first version of the armband (the SenseWear® Pro) and the associated software had algorithms that only distinguished between rest and activity. Algorithms for specific activities such as walking, biking, and stair climbing were available, but the user had to choose the context. For the current armband, the SenseWear® Pro2, the algorithms accurately classify many activities automatically and user selection of an appropriate algorithm is no longer required...
  The poor performance of the algorithms on standing arm swings brings up an
  important question of any measurement technique – what is the range of activities that can be monitored accurately using the device? 

So my summary would be that it's certainly not snake oil, and there is published data showing that it's reasonably accurate in practice (acceptable randoms and systematic errors in most cases), but the actual science behind it is partly proprietary and therefore can't be independently evaluated.
