Lots of people have tried, and continue to try, to "measure the blood pressure" using a non-invasive wearable device. The fact that the there is no significantly adopted wrist-worn blood pressure device on the market, and the category leader Omron has just now announced a wrist blood pressure "watch"--albeit one that still uses inflatable pressure cuff just like their bulkier standard brachial cuff device seen in clinics--should alert us to the difficulty of the endeavor. Part of the problem is that the blood pressure varies a lot--moment to moment--depending on where it is measured. For example, the left and the right arm readings are different, and you just have to search for "wrist and arm blood pressure difference" to find hordes of people complaining about the difference they notice between the more reliable arm reading and the wrist device readings. Nevertheless, companies like Jawbone will still try to make wrist worn blood pressure devices. I first learned about the problem at Jawbone, but this blog is NOT a summary of the Jawbone algorithms; instead, this is a self-study of the signal processing algorithm side of the blood pressure estimation: purely what I can derive by myself.
Terminology
Pressure at any point in the artery has this general pattern:
AP (augmentation pressure) is the reflected pressure wave (green in the above picture), so its phase relative to the main pulse pressure will depend on the distance from the heart. With this waveform, the augmentation index cAI = AP/PP. If the green hump moves farther out to the right, the there will be no AP, like this case:
In this case, we can instead measure the peripheral augmentation index pAI = P2/P1.
When you measure either the bioimpedance or optical signal available at the wrist, you get the raw signal that looks like the above pattern when plotted--that is the relatively straight-forward part; the difficult part is backing out the (central or brachial) blood pressure from the observation at the wrist by relating to the ground-truth pressure (the above patterns)--either calibrated ahead of time for an individual, or established a priori for population group(s).
On indefinite hold
Before I learned about the physics of the problem, I got pulled off to bring up an alternative sensor scheme...
