Why consumer sleep accuracy is a trap (and what to measure instead)

Sleep labs validate against polysomnography. Consumer sleep trackers validate against nothing stable — not other wearables, not consistent behavior across nights, not any external ground truth. The result is a market full of stage charts and sleep scores that agree with each other about 30% of the time, which means they agree with each other about nothing at all.

Chasing PSG accuracy in a consumer product is a trap, not because phones are inaccurate, but because accuracy is the wrong objective. The phone is measuring the wrong things entirely.

What a phone on the nightstand actually measures

Let that sink in: a phone is not a wrist device. Its accelerometers are stationary. Its microphone is an open air path into a room. Its sensor fusion is dominated by acoustics, not motion. A consumer phone sleep product is really an acoustic event detector, and every honest one should start there.

From an acoustic standpoint, the detectable events have stable, reproducible definitions:

• Snoring episodes: an aperiodic broadband burst in the vocal-band envelope with spectral tilt falling off above 1.5 kHz.
• Breathing pauses: a sub-baseline spectral flatness over a 10–30 second window in the 100–400 Hz band.
• Positional shifts: detectable from the tilt sensor over the course of the night.

These are not proxies for sleep stages. They are real events with clinical correlates (snoring intensity correlates with sleep-disordered breathing; breathing pauses correlate with apnea).

The accuracy fallacy

If the goal is "match the lab EEG," a phone will fail — but so will a wrist device, because the correlation between actigraphy and EEG sleep stages is weak by construction. The question is not whether a phone matches a lab. The question is whether the acoustic events it detects correlate with outcomes users care about: morning fatigue, daytime sleepiness, partner-reported disturbance.

That is the metric worth measuring: outcome, not stage.

Measuring against outcomes

SleepTrace measures its own accuracy against a daily morning fatigue slider and a weekly partner-disturbance score — both self-reported. The acoustic event detection (snore duration, breathing pauses, position changes) is validated against those outcomes, not against PSG. The technical writeup of the acoustic feature set that supports this is on the blog.

For builders: stop benchmarking against the wrong bar

Stop optimizing for PSG alignment. Optimize for outcome correlation. Ship the events, ship the confidence, ship the gaps. Let the stage charts be the domain of devices that earn a medical predicate. A phone on a nightstand earns its keep by being honest about what it can hear — and what it cannot.