Why "sleep stage detection" is the wrong metric for a phone app

Sleep apps ship stage charts like they're medical readouts. A pie chart that says "you were 23% deep sleep" feels authoritative, but it is a guess built on movement, breathing rhythm, and time-of-night priors — not an EEG. A phone on the nightstand does not have access to the brain waves that actually define N1, N2, N3, and REM.

That does not make phone-based sleep tracking useless. It makes it useful for the things that are actually measurable and actually matter: when you snored, how often you gasped, whether your breathing paused, and how consistent your rhythm was night to night. Those are acoustic events, not inferred brain states.

The mistake is presenting proxies as measurements. The fix is presenting measurements as proxies.

What a phone can actually hear

The accelerometer on a phone on the nightstand gives you movement: did the phone shake, did the table vibrate. That is coarse. The microphone gives you a far richer signal. You can detect:

• Snoring intensity and duration per sleep cycle.
• Breathing pauses that correlate with apnea events.
• Positional snoring patterns (back vs. side).
• Environmental noise that fragments sleep.

None of that maps to "Stage 2 sleep," but a user who learns "you snored for 18 minutes in the 01:00 cycle and gasped twice" has something actionable. A user who learns "deep sleep dropped 3%" does not, because deep sleep on a phone is a probabilistic model output dressed as fact.

The right metric stack

1. Events, not stages. Lead with concrete detected events: snore episodes, breathing pauses, position changes.
2. Consistency, not snapshots. "Three nights in a row over your baseline" beats "your score is 71/100."
3. Behavioral levers. Tie events to inputs: "on nights with alcohol after 8pm, snoring doubled."
4. Confidence flags. Tell the user when the signal was weak (phone too far, noisy environment) rather than hiding uncertainty behind a number.

We built exactly this in SleepTrace: the iPhone on the nightstand runs an on-device audio pipeline that surfaces events and trends, not a stage pie chart you can't act on. The full technical write-up of the acoustic detection is on the blog.

For builders: stop optimizing for the chart

If you are deciding what to ship, ask not "does our model predict sleep stage" but "does the user change behavior after seeing this number." The latter is honest. The former usually is not.