Why Sleep Studies Disagree: A Field Guide to PSG, Actigraphy, EEG Headbands, and Radar

Two Trackers, Two Answers, One Confused Sleeper

You wear an Apple Watch on your left wrist, an Oura Ring on your right hand, and a Withings mat under your mattress. In the morning, all three tell you a different story. The Watch says you got 7h 12m. Oura says 6h 41m. Withings says 7h 28m. Your "REM" numbers don't even share a decimal place.

Welcome to the swamp. Every consumer sleep tracker has its own answer, and the gold-standard medical test, polysomnography, is so impractical you'll probably never take one. So which device is right? Usually none of them, completely. They're all approximating something they can't directly observe, with sensors that disagree by design.

This post is a field guide. We'll go through every method that matters in 2026, the actual peer-reviewed accuracy numbers vs PSG, what each one is good at, and what each one is dishonest about. No marketing copy. Receipts only.

Key facts

• PSG agreement between human scorers is around 83 percent on a 5-stage epoch-by-epoch basis, so even the gold standard isn't a single number (Younes 2018, JCSM)
• Wrist actigraphy has 96.5 percent sensitivity to sleep but only 32.9 percent specificity to wake (Marino et al, Sleep 2013)
• Consumer wearables as a class show high sleep sensitivity (>0.93) and low specificity (0.18-0.54), meaning they call almost everything sleep (Chinoy et al, Sleep 2021)
• Oura Gen 3 with the OSSA 2.0 algorithm hits ~94 percent sleep sensitivity and ~73 percent wake specificity, the best of the consumer wearables (Altini & Kinnunen 2024)
• Dreem 2 EEG headband scores sleep stages at ~83.5 percent accuracy, in range with individual human scorers (Arnal et al, Sleep 2020)
• AASM guidance: PSG remains the diagnostic standard for OSA in complicated cases; home sleep apnea tests are an alternative for uncomplicated cases at high pre-test probability (Kapur et al, JCSM 2017)

Polysomnography (PSG): The Gold Standard, Explained

PSG is what happens in a sleep lab. You show up, a tech glues 20-plus electrodes to your scalp and face, you sleep in a strange bed, and a sleep physician scores your night in 30-second epochs the next morning. It records EEG (brain waves), EOG (eye movements), EMG (muscle tone), ECG, airflow, chest/abdominal effort, and pulse oximetry simultaneously (AASM Scoring Manual overview).

It's the standard because it's the only method that directly measures the things that define sleep stages. REM sleep is defined by EEG patterns plus eye movements and atonia. You can't infer it from a wrist. You actually have to look at the brain.

But "gold standard" doesn't mean "perfect." Inter-rater agreement between trained human scorers on PSG is roughly 80-83 percent for 5-stage classification (Younes 2018, JCSM). Two qualified humans looking at the same brain will disagree on roughly 1 in 5 epochs. That's the ceiling everything else is being measured against.

PSG is also expensive (often $1,000-$3,000 in the US), inconvenient, and a single night doesn't capture the variability of sleep across weeks. The AASM recommends in-lab PSG for diagnosing obstructive sleep apnea in people with significant cardiopulmonary disease, neuromuscular weakness, or chronic opioid use, and home sleep apnea tests as an alternative for uncomplicated high-risk adults (Kapur et al, JCSM 2017).

Strengths: direct EEG, ground truth for stages, validated for clinical diagnosis. Weaknesses: cost, single night, lab environment, inter-rater variability.

Actigraphy: The Old Reliable

Actigraphy uses a wrist-worn accelerometer to infer sleep from movement. The classic device is the Philips Actiwatch, used in research and recommended by the AASM for tracking circadian rhythm disorders (Smith et al, JCSM 2018).

The biggest validation is Marino et al, who looked at 232,849 epochs and found actigraphy had 96.5 percent sensitivity to sleep, 86.3 percent overall accuracy, but only 32.9 percent specificity to wake (Marino et al, Sleep 2013). Translation: if you're lying still, actigraphy thinks you're asleep, even when you're not.

This bias matters. People with insomnia move less when awake, so actigraphy systematically overestimates their sleep. It's good for population studies and shift work research. It's not good for "did I sleep well last night."

Strengths: cheap, validated, multi-night data, decent for total sleep time in healthy people. Weaknesses: terrible at wake detection, can't stage sleep at all, biased for insomnia and old age.

Consumer Wrist Wearables: Apple Watch, Fitbit, Garmin, Whoop

These are dressed-up actigraphs with extra sensors: optical heart rate (PPG), skin temperature, sometimes SpO2. The PPG signal lets the algorithm guess at sleep stages by reading heart rate variability patterns.

The seminal benchmark is Chinoy et al 2021 in Sleep journal, which tested seven consumer devices simultaneously against PSG. The headline: all of them had sensitivity ≥0.93 to sleep but specificity to wake of just 0.18 to 0.54 (Chinoy et al, Sleep 2021). They all systematically miss your awake-in-bed time. Sleep stage results were "mixed."

A more recent 2025 study tested six newer devices including Apple Watch Series 8 and Fitbit Sense, with PSG runs through August 2024. Same pattern: >90 percent sleep epoch sensitivity, specificity 29-52 percent (Sleep Advances 2025).

For Whoop specifically, validation showed 89 percent agreement on 2-stage (sleep/wake) and just 64 percent agreement on 4-stage classification, with REM sensitivity around 51 percent (Berryhill et al, J Sports Sci 2020). Cohen's kappa of 0.47 is "moderate" agreement, which is academic-speak for "useful direction, not the truth."

Strengths: passive, multi-night, getting better fast, decent total sleep time. Weaknesses: still can't see wake well, sleep staging is statistical guesswork on PPG, no position data, you have to wear them.

Smart Rings: Oura

Same family as wrist wearables, but the finger has cleaner PPG signal because there's less ambient light contamination and motion artifact. It shows in the numbers.

The 2024 multi-night ambulatory validation of Oura Gen 3 with the OSSA 2.0 algorithm is the strongest result for a consumer wearable yet: sensitivity 94.4-94.5 percent, specificity 73.0-74.6 percent for sleep/wake, with 421,045 epochs across 96 participants (Altini & Kinnunen 2024). A 2025 systematic review in OTO Open echoed that Oura now performs in the range expected for cardiorespiratory-based sleep staging (Khan et al 2025).

If you're going to wear something to bed, Oura currently has the best validation data in the consumer space.

Strengths: best-validated wearable PPG, clean signal, comfortable. Weaknesses: still PPG-based staging (not EEG), no position, has to be charged, can't see your environment.

Under-Mattress Mats: Withings, Beddit, Eight Sleep

These use a thin sensor strip under the mattress to detect ballistocardiogram (the tiny vibrations of your heartbeat) and breathing. They sit between contactless and contact: nothing on your body, but something has to be installed.

Validation is best for the Withings Sleep Analyzer for sleep apnea screening. In 118 patients with simultaneous PSG, the device detected moderate-to-severe sleep apnea with 88.0 percent sensitivity and 88.6 percent specificity at AHI ≥15 (Edouard et al, JCSM 2021). That's clinically useful screening, not diagnostic, but it's better than the wrist class for breathing-related events.

Sleep staging on these mats is similar to wrist wearables in quality: total sleep time and broad wake/sleep agreement is okay; REM/NREM stage detection is approximate.

Strengths: nothing to wear, good for apnea screening, captures breathing patterns. Weaknesses: needs install, can be confused by partners or pets, sensitive to mattress thickness, no position info.

EEG Headbands: Dreem, Muse S, Frenz

Now we're back to actually measuring brain activity, just with dry electrodes and fewer channels. These devices are the only consumer-grade products that can plausibly stage sleep, because they record real EEG.

The Dreem 2 headband validation in Sleep journal hit 83.5 percent overall accuracy on 5-stage staging vs PSG, compared to 86.4 percent average among five human experts (Arnal et al, Sleep 2020). That's the closest any consumer device has come to matching individual human scorers.

A 2025 follow-up looked at Dreem in older adults and people with Alzheimer's disease and found total sleep time and sleep efficiency within 10 percent error, but underestimated wake-after-sleep-onset by ~17 minutes (Sleep 2025). 5-stage performance was "moderate" in that population, which tracks with how messy older-adult sleep architecture really is.

Frenz Brainband, a newer single-channel EEG plus actigraphy device, reports similar numbers in their published validation: ~88 percent sleep scoring precision in their Oxford-Colorado studies (Truong et al 2024). Still well short of PSG, but firmly above the wrist class.

Strengths: real EEG, can actually stage sleep, best non-PSG architecture data. Weaknesses: you have to wear something on your head, comfort varies, expensive, often requires phone connection.

Contactless Radar: Where Komori Lives

Radar is the technology Komori has chosen for our contactless wellness monitor. The principle: tiny FMCW (frequency-modulated continuous wave) radar at 60 GHz can resolve millimeter-scale chest-wall movement from across the room. That gives you data about chest-wall motion, gross movement, and body position, all without anything touching you. (Heart-rate and respiratory-rate metrics from radar are a Pro-sensor-suite capability under separate FDA-dialogue scope, deployed under research/IRB protocols — not part of the consumer Komori product.)

Recent academic work has shown high accuracy for the basic vital signs. A 2024 Scientific Reports study using mmWave radar showed respiration rate accuracy >90 percent in static conditions (Nature Sci Reports 2024), and a recent JCSM clinical study showed wireless radar combined with deep learning could classify obstructive sleep apnea severity well enough to be clinically useful (JCSM 2024).

The honest part: radar is not yet validated for REM/NREM staging. Some research groups have published exploratory results, but there is no large multi-night PSG validation that establishes radar staging at the level Oura or Dreem have reached. So Komori doesn't claim to stage sleep. We give you what Komori reliably reports: position, movement, bed exit (more on what Komori does).

Strengths: no wearable, no camera, captures position and movement, environmental data, is being designed to handle couples and pets when configured correctly, local processing for privacy (more on no-camera monitoring). Weaknesses: not staging-validated yet, can't see brain activity, single-source so it can't tell you about HRV-style autonomic data.

How They Stack Up

Numbers are pulled from the validation studies cited in each section. Different studies use different protocols, so don't read decimal points as gospel.

Sensitivity vs Specificity, In Plain English

Most of the disagreement above comes from one tradeoff. Let's make it simple.

Sensitivity: when PSG says you're asleep, does the device also say you're asleep? Almost every consumer tracker is great at this, because anything still that's vaguely breathing looks like sleep.

Specificity: when PSG says you're awake, does the device also say you're awake? This is where everything that isn't EEG or radar fails. Lying still in bed wishing you were asleep looks identical to sleep to a wrist accelerometer.

This is why insomnia studies prefer EEG. It's why your watch's "you slept 8 hours" can be wildly wrong if you spent 90 minutes scrolling at 3 AM without moving. And it's why the right device depends on what question you're asking (more on why we don't show a sleep score).

If you care about whether you're getting enough rest, total sleep time from a good wearable is fine. If you care about why you keep waking up, you want something that can actually see wake. If you care about how often you flip to your back and start snoring, you want position data (more here).

So What Should You Actually Use

Three honest answers:

For diagnosing a sleep disorder: see a sleep specialist. PSG or a doctor-prescribed home sleep apnea test is the right tool. Consumer trackers are not diagnostic. If you suspect apnea, neurologic issues, or your sleep is making your daytime life worse, talk to your doctor.

For tracking your own sleep over time: Oura Ring is currently the best-validated wearable. If you don't want a wearable, Withings Sleep Analyzer screens for apnea and gives you broad sleep/wake data. If you want actual brain-stage data and don't mind a headband, Dreem-class devices are the consumer EEG winners.

For tracking what your body and environment are doing: that's where Komori sits. Radar is the only sensor that captures position, gross movement, and bed exit, contactless, with no camera, locally processed (what we measure). We don't claim REM/NREM. We don't claim vital signs. We don't fake it. We give you the things radar reliably knows.

The dream device is a stack: an Oura on your finger for HRV and broad architecture, a Komori on your nightstand for position and environment, and once a year, a sleep specialist if anything looks off. Different sensors, different jobs, different truths.

That's the field. Pick what answers your question.