Hacker News Reader

Comment by tejaswiy | Hacker News Reader

Since noone's asked this question yet, how accurate is using movement / heart-rate as a proxy for EEG when estimating sleep patterns?

View on HN

Mine didn't have HR monitoring, but I found the movement (aka actigraphy) to be pretty much useless. Fitbit was counting all the time I was laying still trying to fall asleep as "deep sleep", and looking at the chart I couldn't differentiate movement at 12:00 when I knew I was awake from movement at 6:00 when a noise woke me up but I fell right back asleep.

The movement data is probably enough to tell when I've slept better or worse, but I wouldn't trust the absolute "number of hours of sleep" they report.

Fitbit does not attempt to track sleep stages w/o HR monitoring; it's just Awake, Restless or Asleep. Yes with nothing but accelerometer data stillness can be mistaken for Asleep, but I think you're misremembering when you say "deep sleep".

It depends on the device. I used to use the Microsoft Band 2, and it seemed pretty good to me - at least it helped to validate what I believed to be my sleep patterns. In contrast, I now use a Garmin Fenix 3, and it's completely and utterly worthless for sleep tracking - it routinely decides I'm sleeping when I'm in bed even if I'm on my phone or change the display on the watch, and it rarely picks up the times I wake up during the night. --It's too bad, because for activity tracking throughout the day it's actually really good.

But doctors don't seem to trust the consumer devices (probably with good reason). When I had an actigraphy, they used a special device made by Phillips, and used the results to verify that the sleep study results weren't a one-off night.

My own research has shown that a single EEG channel with traditional techniques gets you into the ballpark of 60% accuracy (another human rater on all sorts of data gets you to ~80% accuracy and state of the art methods around 75% accuracy on a single EEG channel). Movement + heart-rate has much less information, so I'd expect about 40% accuracy overall.

My work has been traditionally done with AASM sleep labels which may be a bit more complex that what fitbit uses, so that may bring up their overall accuracy. Additionally I assume that fitbit would use a hidden markov model to predict the sleep stages (which can produce very deceptive results in the context of sleep IMO). I'd say that this would bring their overall accuracy up to around 55% based upon my best guess.

And are there any good sleep trackers?

Zeo made a very good one called something like Zeo Sleep Manager. They stopped making them a few years ago, but you can still find them online and there's a fairly active community around them still.

Yes, and the Zeo is a REAL sleep tracker in the sense that it uses EEG -- not accelerometers -- as sensors (where the latter actually don't measure anything that has to do with sleep, just a proxy that is somewhat correlated but also easily fooled). Remember, when you get a real lab sleep study done, the primary measurement of sleep is via the spectral powers of your brain waves via EEG -- these measurements represent your sleep stages, by definition.

Reply on news.ycombinator.com