How accurate are consumer wearables? A look at validation studies

Consumer wearables have improved markedly in the last decade, but they are not lab equipment. Knowing where they're accurate and where they're approximations changes how you use the data they produce.

Heart rate (resting)#

Validation studies (Bent et al. 2020, Düking et al. 2020) consistently show:

• Wrist-based photoplethysmography (PPG) agrees closely with chest-strap ECG at rest, typically within 1–3 bpm.
• Ring-based PPG (Oura) shows similar accuracy at rest.
• Skin tone affects PPG signal quality; darker skin produces higher error rates in older optical sensors. Modern devices have closed much but not all of this gap.

For tracking resting heart rate trends, consumer devices are reliable.

Heart rate (exercise)#

Accuracy drops during exercise:

• Steady-state aerobic exercise: typically within 5–10 bpm.
• Variable-intensity exercise (intervals, weight training): much larger errors, particularly during eccentric movements and weighted wrist work.
• Cold weather, sweat, and loose fit increase error.

For accurate HR during hard training, a chest strap is still the consumer-accessible gold standard.

HRV#

Validation versus ECG (Stone et al. 2021; Bellenger et al. 2021):

• Sleep-time HRV from modern wearables shows reasonable correlation with ECG.
• Awake HRV from wrist devices has much larger error and is generally not recommended for serious training-decision use.
• Different devices report different transformations of RMSSD; absolute values are not comparable across brands.

This is why most wearables report HRV only during sleep. It's where the signal-to-noise ratio is acceptable for consumer hardware.

Sleep duration#

PSG-comparison studies (de Zambotti et al. 2017–2024 series):

• Total sleep time: most modern devices within 10–30 minutes of PSG.
• Sleep efficiency: similar agreement.
• Detection of sleep onset: tends to misclassify quiet wakefulness as sleep, leading to slight underestimation of latency.

For tracking how long you slept and how broken the night was, modern wearables are reasonable.

Sleep stages#

This is where consumer wearables are weakest:

• Four-stage classification (wake/light/deep/REM) accuracy versus PSG is typically 50–70% of epochs correct.
• REM and deep are most often confused with each other.
• Single-night absolute values for deep sleep or REM should not be taken literally.
• Within-individual trends over weeks are more reliable than absolute values.

If a wearable tells you "you got 47 minutes of deep sleep last night", treat that number with substantial scepticism. If it tells you "your deep sleep has been trending lower for a week", that trend is more credible.

SpO2#

Wearables include SpO2 estimates with disclaimers; the disclaimers are warranted. Accuracy versus medical-grade pulse oximeters varies:

• Healthy adults at rest: typically within 2–4%.
• Sleep apnoea screening: not a substitute for proper diagnostic testing.
• Detection of clinically significant desaturation: poor in some independent studies.

Wearable SpO2 is not a diagnostic tool. If you suspect sleep apnoea, your GP is the right route to a proper sleep study.

Skin temperature#

Reasonable correlation with body temperature for trend purposes. Useful for detecting illness onset (small upward shifts can precede symptoms by 1–2 days). Cycle-tracking based on temperature shows reasonable evidence.

What to take from this#

• Treat absolute single-night numbers with appropriate scepticism, particularly stage classifications.
• Trust trends more than points. Week-on-week direction is where wearables shine.
• If you have a clinical question (apnoea, insomnia, arrhythmia), a wearable is not a diagnostic tool. See a GP and get proper testing.
• Pick a device, stick with it, and stop comparing your numbers to anyone else's.

The wearable category is at the point where the most-validated metrics (total sleep time, resting HR, sleep HRV trends) are useful. The least-validated metrics (exact sleep stage minutes, wearable SpO2 for sleep apnoea, awake HRV) are over-marketed relative to what the validation literature supports.