New non-stages sleep algorithm overestimates sleep more than old algorithm

[bluefoxicy]

This post is about the accuracy of the current non-stages sleep algorithm, not the UI and feature set changes (we have a large thread about that).

The new sleep algorithm used for Fitbit One, Fitbit Charge HR, and other non-stages devices overestimates sleep more than the prior algorithm.  Fitbit Premium still uses the prior algorithm, which has more-reliable wake-event detection.

A comparison below:

Old and new algorithm side by side, with 3am-5am WASO

The above event has a large wake-after-sleep-onset (WASO) between 3am and 5am.  The algorithm used on Fitbit Premium correctly-identifies this period as largely awake.  The new algorithm detects this segment as largely restless.  Both read 8h14m total sleep time (TST).

Potential Improvements

My Charge HR actually automatically detected sleep beginning at 11:28 and ending around 3:30am, with a second segment starting 5:17am.  As well, the Fitbit reads the time to fall asleep or sleep latency (SL) as 3 minutes in this case; the correct SL is 20 minutes, with sleep beginning at 11:28.  This has some implications.

The industry-standard for detecting sleep onset is 20 consecutive epochs with no more than one epoch detected as awake.  Using that metric, the period from 11:11 to 11:21 on 30-second epochs is 20 epochs.  It measures two awake epochs at 11:19, near as I can tell. On one-minute epochs, the definite onset of sleep is 11:28; it appears to be 11:28 on 30-second epochs as well.

If the Fitbit detects a WASO event in a solid sleep segment, it currently ends that segment until a new sleep event begins.  If the user measures it as one segment (because Premium doesn't give you a total sleep efficiency if you have multi-segment sleep), Fitbit doesn't record it as WASO, but rather as a period with a lot of awake epochs.  There isn't a compromise to record significant WASO events as awake by using the same logic as above to determine when sleep begins again.

[LanuzaFitbit]

Hello @bluefoxicy! 

Thank you very much for your in-depth explanation and details. I have to say that I truly enjoyed the read. 

That is certainly a very interesting thing to point out.Be sure that no comment goes unnoticed. I will be passing this on to our team.

Thank you very much for your feedback.