Recording Wearable and Polysomnography Side by Side
One night, two sensors, one sleeping body under observation.
- 12+: PSG channels recorded (EEG, EOG, EMG, ECG)
- 2–3: Wearable sensors (accelerometer, PPG, temp)
- 30 sec: PSG epoch length (AASM scoring standard)
- 30–60 sec: Wearable epoch length (algorithm dependent)
Polysomnography is the reference method
PSG scores sleep from brainwaves, eye movement, and muscle tone.
Wearables infer sleep indirectly
Wrist devices guess sleep stage from motion and heart rate only.
No consumer wearable measures brain activity directly tonight.
Same night, two very different data streams
Both devices timestamp epochs so their outputs can be aligned.
Total Sleep Time — Wearables Run a Bit Long
Quiet stillness in bed often gets counted as sleep by mistake.
- 10–30 min: Typical TST overestimate (per night, basic devices)
- ±10 min: Sleep onset error (vs PSG onset latency)
- high: Wake-after-sleep-onset miss rate (quiet wake looks like sleep)
- ~50%: Advanced algorithm improvement (lower TST bias)
Stillness is mistaken for sleep
Lying motionless while awake reads the same as light sleep.
Sleep onset is usually detected late
Devices wait for sustained stillness before declaring sleep onset.
Total sleep time bias grows on more fragmented nights.
Better algorithms shrink the gap
Adding heart-rate variability narrows the wearable overestimate.
Light vs Deep Sleep — The Hardest Call
Motion and heart rate barely differ between light and deep sleep.
- ~50–60%: Light/deep agreement (basic) (vs PSG staging)
- ~70–80%: Light/deep agreement (advanced) (with PPG + HRV)
- common: Deep sleep underestimation (short N3 bouts missed)
- blurs: Epoch smoothing effect (short stage transitions)
Deep sleep has few visible signals
N3 sleep looks physiologically similar to light sleep on wrist sensors.
Short bouts get smoothed away
Brief deep-sleep epochs are often merged into light sleep.
Movement artifacts push more epochs into the wrong bucket.
Machine-learning models help, but imperfectly
Modern classifiers raise agreement without matching EEG precision.
REM Sleep — Detectable by Heart Rate Rhythm
REM sleep raises heart-rate variability enough for wearables to notice.
- ~60–65%: REM detection accuracy (basic) (vs PSG REM epochs)
- ~80–85%: REM detection accuracy (advanced) (PPG-based algorithms)
- 4–6: REM cycles per night (roughly 90-min spacing)
- increases: False REM in disrupted sleep (with more arousals)
Heart rate rises and steadies in REM
Autonomic signals give wearables a usable REM fingerprint.
Motion adds little extra help
Muscle atonia in REM means near-zero movement, like deep sleep.
REM and deep sleep can look alike without brainwave data.
Disruption confuses REM timing
Frequent arousals scatter false REM flags across the night.
Accuracy Summary — Trends Yes, Precision No
Wearables track sleep patterns well but miss exact stage boundaries.
- ~65–85%: Overall epoch agreement (algorithm dependent)
- Trends: Best use case (week-to-week consistency)
- Diagnosis: Weakest use case (clinical staging precision)
- Gold standard: PSG remains (for medical sleep studies)
Useful for longitudinal self-tracking
Night-to-night trend direction is usually reliable enough to act on.
Not a diagnostic substitute
Clinical sleep disorders still require lab-grade polysomnography.
Treat wearable sleep stages as an estimate, not a diagnosis.
The gap is narrowing over time
Newer sensors and algorithms keep closing the accuracy gap.