Skip to main content

Wearable Sensors and Measurements

Motion & Step Counts​

Physical phenomenon
  -> Sensor
  -> digital sampling
  -> time indexing
  -> quality checking
  -> filtering
  -> windowing
  -> feature extraction
  -> modeling
  -> metric display
  -> API record

The key thing is that each arrow here encodes a choice that changes what is recorded.

Think of the Inertial Measurement Unit (IMU) as a box with springs (at least for the acceleration bit, it does angular momentum calcuations as well.) It’s a six-axis device: six signals.

What do you do with six signals? You can compute Magnitude and do a bunch of things with it. Note that traditional approaches to things like Step Counting may suck because e.g. walking with a cart (wrist stable) or riding a bus (lots of bumping up and down.)

Deep Learning methods might be cooler. TODO: Are they retraining the Neural Nets per individual?

ENMO is “Euclidean Norm Minus One” which is x2+y2+z2−1g\sqrt{x^2 + y^2 + z^2} - 1g (Note that the g is gravity!)

Heart Rate​

Photoplethysmography: use Green Light to reflect and Red to absorb. In general. TODO: look at the doublecurve; that’s backflow? there’s names for each peak.

High HRV (heart rate variability) means you might be stressed, for example.

note

All we are measuring is physiological stress and using physiological signals.

Note that ECG captures electrical activity while the PPG does not.

Note that Norm-to-Norm is for ECG data. Inter-beat Interval is for PPG. It’s a kinda naming thing.

Sleep​

Polysomnography (PSG) is the gold standard. LOTS of measurements! You use PSG data to assign sleep labels. You can compare these to PPG (and find that the latter is more of a rough estimate.) You certainly can use multimodal models (motion, HRV, SpO2, skin temp.)

Digital Phenotyping​

What kind of sleeper are you? What kind of ‘awaker’ are you (owl or lark)?

QC​

This comes before any kind of features. Data, data, data. The idea here is: what do you do with a continuous signal to turn it into some dataset that your algo/model can crunch on?

Note that different modalities can inform each other.