OxWearables/biobankAccelerometerAnalysis

Overclassification of sleep for low frequency aquisitions (5Hz)

Closed this issue · 3 comments

SuperCorks commented

Hey guys, let me start off by saying thank you very much for this library! If I can tip you in anyway, please let me know how.

We're developing our own device for a study here and gathered a first dataset. The wearable doesn't have too much memory (15MB) and the motion sensor (accel/gyro) is the sensor that produces the most data (12 bytes per sample) so we limited it at 5Hz. Also the sensor can be configured with a precision/max value tradeoff so we went with a max value of 2g (other options are 4g or 8g).

We're now running different actigraphy algorithms, including yours, on our data and seeing some weird results as shown below:

newplot1
newplot2
image

I understand our setup isn't ideal, so here are my questions:

  • It looks like the model over-identifies sleep. With our current 5Hz setup, is there something we can do to make the models more accurate?
  • What is the best configuration of our sensor that your suggest? Here are some possible examples:
    • 20 Hz with a max of 4g?
    • 32 Hz with a max of 4g?
    • 32 Hz with a max of 8g?
angerhang commented

Thanks for your questions. It is always nice to have external validation using our algorithm. 5Hz does seem a bit too low for sleep, making the device less sensitive to movement. To me 4/8g in terms of range doesn't make a big difference but something with 30Hz would likely help with your classification output.

@chanshing @R-Walmsley @aidendoherty do they also need to normalise the input for the random forest model by any chance?

chanshing commented

5Hz is indeed way too low. For example, slow walking is at around 2Hz already (2 steps/sec), meaning you need at least >4Hz to capture it (Nyquist criteria). Sprinting is almost 5Hz, so >10Hz there.
In terms of range, 4g is more than enough from my experience (I remember something like <0.1% of samples of free-living data exceeds 3g).
For reference, Fitbit and Apple Watch use something like >100Hz and +/-16g which I think might be too much. I think +/-4g @ 50Hz should be enough for most cases. Given your options, I would go for the 32Hz, 4g

SuperCorks commented

Alright thanks to all of you for you answers!

  • 👍1