What Most People Think
The number on your wrist after a workout feels like a biological fact. 487 calories burned — precise, specific, trustworthy. The sleep breakdown reads like a lab report: 1 hour 12 minutes of deep sleep, 1 hour 34 minutes of REM. The stress score feels like a real-time readout of your autonomic nervous system.
Apple, Garmin, WHOOP, and Oura spend billions on marketing that reinforces this perception. The interfaces are clinical. The numbers look exact. So people build their nutrition, recovery, and entire training programs around them — trusting that these proxy measurements carry the same weight as the clinical gold standards they're mimicking.
The Practical Takeaway
- Trust completely: Resting heart rate trends, overnight HRV trends, total sleep duration, and sleep/wake detection. These are clinically useful for long-term lifestyle monitoring.
- Ignore entirely: Workout calorie counts for nutritional planning. Never eat back "burned" calories. A reported 500 kcal burn may really be 250–400 kcal.
- Use directionally: Sleep staging as week-over-week trends only — not single-night absolutes. "My deep sleep has been declining for 3 weeks" is useful. "I got 47 minutes of deep sleep last night" is noise.
- Screen, don't diagnose: OSA detection is good enough to flag a problem — then get a real sleep study.
- Avoid orthosomnia: Obsessing over inaccurate sleep data can worsen actual sleep quality via nocebo effect. If you feel rested, your device's opinion is irrelevant.
Quick Answer
Your smartwatch is a legitimate medical instrument for exactly two things: resting heart rate and knowing when you're asleep. Everything else — calorie burn (20-90% error rates), sleep stage breakdowns (up to 43 minutes wrong on deep sleep), and real-time stress scores — ranges from "directional guess" to "fiction."
The calorie number after your workout is the most dangerous metric on your wrist: it's consistently wrong enough to sabotage a cut if you eat based on it.
What the Evidence Shows
Researchers have tested every major consumer wearable against the clinical gold standards: indirect calorimetry for calories, polysomnography for sleep, and electrocardiography for heart metrics. The results split cleanly into "surprisingly good" and "embarrassingly bad."
Heart rate at rest is genuinely accurate. Across 60 participants wearing 7 different devices, wrist-worn HR monitors achieved greater than 95% agreement with clinical ECG during rest and steady-state exercise (Shcherbina et al., 2017). HIGH Apple Watch achieved the lowest error. During military close-quarter battle training with 50 special forces soldiers, Apple Watch hit 0.6% mean absolute percentage error for HR — essentially clinical grade (Alvarez-Garcia et al., 2024).
What would change this: Evidence of systematic HR drift during prolonged continuous monitoring (>72 hours) or failure in populations with atrial fibrillation.
Calorie burn is not remotely accurate. In the Shcherbina study, no device achieved energy expenditure error below 20%. The worst performer hit 92.6% error. A separate validation found Apple Watch overestimating walking calories by 19.8% and running calories by 24.4% — and these were steady-state activities, the easiest scenario for the algorithms (Wang et al., 2022, N=20). LOW
What would change this: Integration of transdermal metabolite sensors (lactate, glucose) or validated open-source algorithms tested across N>10,000 with doubly labeled water.
Sleep detection works — your watch knows when you're asleep. Apple Watch, Oura Ring, and Fitbit all achieved at least 95% sensitivity for detecting sleep versus wakefulness (Chinoy et al., 2022, N=35). HIGH But granular staging is a different story entirely. Apple Watch underestimated deep sleep by 43 minutes in a single night compared to polysomnography. Fitbit Sense correctly identified REM sleep only 60.41% of the time. LOW
Overnight HRV is clinically useful. Resting and overnight heart rate variability measurements show strong concordance with medical-grade ECG (Kim et al., 2025). For tracking autonomic balance, recovery trends, and long-term stress patterns, this data is reliable enough to guide real decisions. HIGH
Trust These
Resting HR (>95% ECG agreement), Sleep/wake detection (≥95% sensitivity), Overnight HRV, OSA screening (92% Sn/Sp)
Don't Trust These
Calorie burn (20-93% error), Sleep staging (up to 43 min wrong), Real-time exercise HR during lifting, Single-night sleep reports
Real World vs Lab
Reality Check #1: The Algorithm Update Problem
Reality Check #2: Body Composition Distortion
Reality Check #3: Resistance Training Destroys Accuracy
Conviction
(Metric-dependent — ranges from HIGH to LOW)
HIGH Conviction
Resting heart rate, overnight HRV, sleep/wake detection, OSA screening
LOW Conviction
Calorie burn, sleep stage breakdown, real-time exercise HR during lifting
What would change this overall assessment: Open-source algorithm validation across anatomically and ethnically diverse populations (N>10,000) using doubly labeled water for energy expenditure and longitudinal home EEG for sleep staging. Integration of transdermal metabolite sensors (continuous lactate, glucose) rather than exclusive reliance on PPG and accelerometry.
What would change my mind on calorie burn accuracy
Devices would need to integrate direct metabolic measurement — transdermal continuous lactate or glucose sensors — rather than inferring energy expenditure from heart rate and movement alone. Open-source algorithms tested across N>10,000 with doubly labeled water validation showing MAPE consistently <10% across all activity types and body compositions.
What would change my mind on sleep staging accuracy
Longitudinal at-home EEG validation (not single-night lab studies) showing wearable deep/REM classification achieving >85% sensitivity AND specificity vs PSG across 30+ consecutive nights in diverse populations. Current devices achieve high sensitivity but poor specificity — they detect sleep but can't reliably tell you what kind.
Sources
- Shcherbina et al., 2017 — Journal of Personalized Medicine, N=60. Seven devices tested against indirect calorimetry and continuous ECG. No device achieved energy expenditure error below 20%. Apple Watch had lowest HR error.
- Wang et al., 2022 — Frontiers in Physiology, N=20. Apple Watch 6 and Garmin FENIX 6 vs COSMED K5 metabolic cart. EE MAPE 19.8–32.0% during walking and running.
- Chinoy et al., 2022 — Nature and Science of Sleep, N=35. Apple Watch 8, Oura Ring Gen3, Fitbit Sense 2 vs polysomnography. Sleep sensitivity ≥95%; deep sleep underestimated by 15–43 minutes.
- Koerber et al., 2022 — American College of Cardiology Systematic Review, N=469 across 10 studies. Skin tone bias: 4/10 studies showed significant reduction, 4/10 found no difference.
- Kim et al., 2025 — Nature and Science of Sleep, N=90. OSA detection via smartwatch AI: 92.3% sensitivity, 92.6% specificity for AHI ≥15.
- Alvarez-Garcia et al., 2024 — European Journal of Sport Science, N=50. Military CQB training validation. Apple Watch HR MAPE 0.6%; both devices failed equivalence testing for energy expenditure.
The Debate
Does Skin Tone Affect Accuracy?
The Nuance
The accuracy hierarchy is device-specific. Apple Watch consistently outperforms competitors in both HR and EE accuracy, though even Apple fails the clinical bar for energy expenditure. If you're choosing a device for health monitoring, brand matters — but no brand solves the fundamental limitation of wrist-based optical sensing for calorie estimation.
Tattoos can completely block green LED penetration, rendering HR and HRV data from that wrist useless. If you have wrist tattoos, wear the device on the other arm.
The biggest irony: wearables are getting genuinely good at detecting serious pathology — sleep apnea screening achieves 92% sensitivity/specificity for moderate-to-severe cases — while remaining terrible at the consumer metric people care most about: how many calories they burned.