Measurement versus Inference in Sleep Assessment: Polysomnography and the Oura Ring Compared
A channel-by-channel comparison of laboratory polysomnography and peripheral ring signals, with a close reading of sleep–wake and sleep-stage validation.
Existing Zenodo DOI: 10.5281/zenodo.23005116
Summary
Sleep stages are defined by cortical electrophysiology. Under the rules of the American Academy of Sleep Medicine (AASM), wake, N1, N2, N3 and REM are scored from electroencephalography (EEG), electrooculography (EOG) and chin electromyography (EMG) [1]. The Oura Ring records none of these signals. It records a peripheral pulse waveform, finger motion, finger skin temperature and reflectance oximetry, and it maps these signals onto sleep stages with a proprietary machine-learning classifier [2,6]. Its sleep stages are therefore inferred, not measured.
This paper compares the two systems signal by signal and output by output, examines what the widely publicised "95% accuracy" figure actually represents, and assesses the independence of the evidence base. Three conclusions follow:
Of roughly twelve signal classes recorded in AASM polysomnography (PSG), the Oura Ring directly measures one (oxygen saturation, by reflectance) and approximates one (cardiac rhythm, via the pulse wave). The remainder are absent or statistically reconstructed.
The 95% figure refers to binary sleep–wake classification and is inflated by base rate. The clinically relevant quantity, wake specificity, is approximately 73%: about one in four epochs of genuine wakefulness is scored as sleep [4].
The studies reporting favourable four-stage agreement were authored, funded or advised by Oura. The only independent study in a clinical sleep-laboratory population found four-stage agreement of approximately 53% [5,7].
1. What Polysomnography Measures
Attended, in-laboratory PSG is the reference standard for sleep assessment. It records cortical, ocular, muscular, cardiac, respiratory and positional signals simultaneously, and a trained technologist scores each 30-second epoch against the AASM manual [1]. Even with the full montage, scoring is not perfect: agreement between accredited human scorers is approximately 83% for stage assignment in healthy adults, and lower in patients with sleep disorders [8]. This figure is the practical ceiling against which any surrogate must be judged.
Critically, PSG does not merely stage sleep. It detects cortical arousals, scores apnoeas and hypopnoeas, distinguishes obstructive from central events, identifies respiratory effort-related arousals (RERAs), and records periodic limb movements, REM sleep without atonia and cardiac arrhythmia. Several of these outputs depend jointly on more than one channel. A hypopnoea, for example, requires a reduction in airflow accompanied by either an oxygen desaturation or an EEG arousal [1].
2. What the Oura Ring Measures
The Oura Ring 4 contains four sensing elements [6]:
Photoplethysmography (PPG): an 18-path, multi-wavelength array. Green and infrared LEDs are used for heart rate and heart rate variability (HRV); red and infrared LEDs for oxygen saturation during sleep. Respiratory rate is derived from modulation of the PPG waveform.
Three-axis accelerometer: finger movement.
Digital temperature sensor: peripheral skin temperature, reported as deviation from the individual baseline.
Oura's staging algorithm combines movement, temperature, heart rate and HRV [2]. Its own foundational validation paper acknowledges that, under AASM rules, sleep staging must be based on and validated against EEG, EOG and EMG [2]. The device therefore estimates cortical states from their autonomic and motor correlates. Those correlates are real: heart rate falls and HRV rises in N3, and both become more variable in REM. But they are indirect, they vary between individuals, and they are disrupted by precisely the conditions that bring patients to a sleep laboratory, including apnoeas, arousals, medication and cardiovascular comorbidity.
One further point deserves emphasis. A PPG pulse interval is not an R–R interval. It includes pulse transit time, which varies with peripheral vascular tone, and vascular tone itself changes with sleep stage and with post-apnoeic sympathetic surges. The ring's cardiac input is therefore confounded by some of the very states it is attempting to classify.
3. Channel-by-Channel Comparison
Table 1. Signals recorded by AASM polysomnography and their Oura Ring equivalents.
| PSG channel (AASM) | What it establishes | Oura equivalent | Status |
|---|---|---|---|
| EEG (frontal, central, occipital) | Defines wake, N1, N2, N3 and REM: alpha rhythm, spindles, K-complexes, slow waves | None | Absent |
| EOG (two channels) | REM (rapid eye movements); sleep onset (slow rolling eye movements) | None | Absent |
| Chin EMG | REM atonia; REM sleep without atonia | None | Absent |
| ECG | Rhythm, arrhythmia, R–R intervals | Finger PPG pulse interval | Surrogate (haemodynamic, not electrical) |
| Oronasal thermal airflow | Apnoea detection | None | Absent |
| Nasal pressure | Hypopnoea; flow limitation; RERAs | None | Absent |
| Respiratory inductance plethysmography (thorax and abdomen) | Respiratory effort; obstructive versus central events | Respiratory rate derived from PPG modulation | Inferred; no effort signal |
| Pulse oximetry | Desaturation; oxygen desaturation index; hypoxic burden | Red/infrared reflectance PPG | Present (reflectance surrogate) |
| Anterior tibialis EMG | Periodic limb movements | Finger accelerometer | Absent (wrong limb, wrong signal) |
| Body position sensor | Positional sleep apnoea | None usable | Absent |
| Snore microphone; synchronised video | Snoring; parasomnia; seizure | None | Absent |
| Scoring | Accredited technologist to AASM rules; inter-scorer agreement ≈83% [8] | Proprietary classifier on HR, HRV, motion and temperature | Unauditable |
4. Output-by-Output Comparison
Table 2. Clinical outputs of polysomnography and the corresponding Oura Ring capability.
| Clinical output | PSG | Oura Ring |
|---|---|---|
| Sleep stages | Measured (cortical activity) | Inferred from autonomic and motor correlates |
| Cortical arousals (≥3 s EEG frequency shift) | Measured | Cannot be detected. Arousals are defined by EEG. |
| Arousal index; sleep fragmentation | Measured | Not available |
| Apnoea–hypopnoea index (AHI) | Measured to AASM 3% or 4% criteria | Not available. The "Breathing Disturbance Index" is not an AHI: hypopnoea scoring requires airflow and either desaturation or EEG arousal. |
| Obstructive versus central apnoea | Distinguished by effort channels | Not possible |
| RERAs; upper airway resistance syndrome | Measured (nasal pressure and EEG) | Not possible |
| Periodic limb movements of sleep | Measured | Not possible |
| REM sleep behaviour disorder | Measured (EMG and video) | Not possible |
| Sleep-onset REM (narcolepsy work-up) | Measured | Not possible |
| Nocturnal arrhythmia | Measured (ECG) | Not reliably assessable from PPG |
| Total sleep time; sleep efficiency | Measured | Estimated; systematically overestimated when wake is misclassified as sleep |
5. The "95% Accuracy" Claim
5.1 What the figure refers to
Oura has marketed "95% sleep staging accuracy compared to clinical sleep lab". In response to litigation, the company has clarified that this figure refers to sleep-versus-wake classification. For four-stage classification (wake, light, deep, REM), it cites agreement of approximately 76–79% [7].
5.2 Why binary accuracy is misleading
Epoch-by-epoch accuracy for a two-class problem is dominated by the more prevalent class. In a healthy sleeper, approximately 85–90% of 30-second epochs in bed are sleep. A device that scored every epoch as sleep, and so detected nothing, would therefore achieve 85–90% "accuracy". A figure of 95% is only modestly better than this trivial baseline.
The informative statistic is specificity: the proportion of true wake epochs correctly identified as wake. In the largest Oura-supported validation (96 participants, 421,045 epochs), sensitivity for sleep was 94.5% but specificity was 73.0% [4].
5.3 Worked example
Consider an eight-hour night with 60 minutes of true wakefulness. At a specificity of 73%, approximately 16 of those 60 minutes are scored as sleep. The resulting error is not a symmetrical 5%. It is a directional bias that inflates total sleep time and sleep efficiency and understates wake after sleep onset (WASO).
This matters because the only non-trivial use of sleep–wake detection is the quantification of wakefulness. No individual requires a device to inform them that they slept. In insomnia, WASO and sleep efficiency are the primary outcome measures, and they are exactly the quantities the device is structurally biased to misreport.
6. Independence and Replication of the Evidence
The studies that report favourable agreement between the Oura Ring and PSG share a common feature: each has a financial or advisory connection to Oura.
Table 3. Principal validation studies of Oura Ring (Gen3) sleep staging against polysomnography.
| Study | Population | Principal finding | Relationship to Oura |
|---|---|---|---|
| Altini & Kinnunen, Sensors 2021 [2] | 106 participants; 440 nights | 79% four-stage agreement | Authored by Oura scientists (algorithm development) |
| Svensson et al., Sleep Med 2024 [4] | 96 generally healthy Japanese adults, 20–70 y; ambulatory PSG | Sleep sensitivity 94.5%; wake specificity 73.0% | Authors report financial support from Oura Health Oy |
| Robbins et al., Sensors 2024 [3] | 35 healthy adults, 20–50 y, no sleep disorder; single night | Four-stage κ = 0.65 | Funded by Oura; lead author on Oura Medical Advisory Board |
| Herberger et al., Sci Rep 2025 [5] | Sleep-laboratory patients | Sleep–wake ≈85%; four-stage ≈53% | Independent |
Two observations follow. First, the favourable results were obtained almost entirely in healthy, carefully selected cohorts, often over a single night. Second, the one independent evaluation in patients referred to a sleep laboratory, the population for whom accurate staging has clinical consequence, found four-stage agreement of approximately 53%. That result is well below the 76–79% cited by Oura and far below the 83% inter-scorer ceiling. Oura has disputed the study's interpretation [7]; it has not been superseded by an independent study showing otherwise.
The defensible formulation is therefore: Oura's four-stage sleep staging performance has not been independently replicated in patients with sleep disorders, and the only such study found approximately 53% agreement with PSG.
All studies cited evaluated the Gen3 ring. The Oura Ring 4 uses revised sensing hardware, and its staging performance against PSG should not be assumed from Gen3 results.
7. What the Oura Ring Legitimately Offers
None of the foregoing implies that the device is without value. It provides continuous, low-burden, longitudinal data on nocturnal heart rate, HRV, peripheral temperature deviation and oxygen saturation. These are genuine physiological signals, and within-person trends in them over weeks and months may be informative for recovery, illness onset and circadian regularity. The error lies in presenting a classifier's output as a hypnogram, and in describing inference as measurement.
8. Conclusion
The Oura Ring does not measure sleep. It measures pulse, finger motion, finger temperature and reflectance oximetry, and it applies a proprietary classifier that maps those peripheral signals onto categories defined entirely by cortical electrophysiology, none of which it records. The mapping is weakest where accuracy matters most: in detecting wakefulness, in disordered sleep, and in every diagnostic output that depends on EEG arousal, airflow or respiratory effort. These outputs it cannot produce at all. Claims of equivalence to polysomnography should be read in that light.
References
[1] Troester MM, Quan SF, Berry RB, et al. The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology and Technical Specifications. Version 3. Darien, IL: American Academy of Sleep Medicine; 2023.
[2] Altini M, Kinnunen H. The promise of sleep: a multi-sensor approach for accurate sleep stage detection using the Oura Ring. Sensors. 2021;21(13):4302.
[3] Robbins R, Weaver MD, Sullivan JP, et al. Accuracy of three commercial wearable devices for sleep tracking in healthy adults. Sensors. 2024;24(20):6532.
[4] Svensson T, Madhawa K, Nt H, Chung UI, Svensson AK. Validity and reliability of the Oura Ring Generation 3 (Gen3) with Oura Sleep Staging Algorithm 2.0 (OSSA 2.0) when compared to multi-night ambulatory polysomnography: a validation study of 96 participants and 421,045 epochs. Sleep Med. 2024;115:251–263.
[5] Herberger S, Aurnhammer C, Bauerfeind S, et al. Performance of wearable finger ring trackers for diagnostic sleep measurement in the clinical context. Sci Rep. 2025;15(1):9461.
[6] Oura Health Oy. Smart Sensing: accuracy and personalisation in Oura Ring 4; and Oura Ring 4 product specifications. ouraring.com; 2024–2026.
[7] What the Oura lawsuit reveals about "accuracy" in consumer sleep tracking. Sleep Review. September 2026. See also: Oura Health Oy. Sleep staging is not guesswork. ouraring.com; 2026.
[8] Rosenberg RS, Van Hout S. The American Academy of Sleep Medicine inter-scorer reliability program: sleep stage scoring. J Clin Sleep Med. 2013;9(1):81–87.