Measurement and validation
In this investigation
The question
We want to improve sleep performance and understand what changed. That requires distinguishing three achievements: a better device grade, a better measured sleep characteristic, and a better health or performance outcome. Evidence for one does not automatically establish the others.
The actual account reports Oura Ring 5, hardware or5, firmware 2.1.25, plus a COROS PACE 4. A supplementary API request returned Oura sleep algorithm version v2; the exact model revision and mapping to validation protocols remain unverified. COROS firmware/algorithm versions were not returned. They remain part of the measurement specification to establish. First record.
As of October 13, 2026, this research has not verified independent peer-reviewed validation of the exact Ring 5 or PACE 4 hardware for sleep staging. That is a limitation of the available evidence, not proof that either device is inaccurate. Published findings from predecessors remain labeled by the hardware and algorithm actually tested.
Quiet waking and extra sleep detections
The October 26 follow-up adds another disagreement: COROS detected 24 minutes of nap sleep in 23:30–00:01, while I explicitly report no naps. Oura separately labeled 19:48–21:04 as rest, with 27m30s estimated asleep; that different window is not confirmed as actual sleep either. Preserve the native classifications and the contradictory diary separately. This is not agreement on a nap or a measured overall error rate. Oura 85/COROS 94 also coexist with reported bad sleep, waking sleepiness, ongoing body pains and feeling unhappy overall; improved grades do not establish the experienced outcome.
Oura infers sleep from peripheral signals rather than measuring EEG. Its own support guidance acknowledges that very still wakefulness can resemble sleep in movement, heart rate and average body temperature. That supplies a plausible error mechanism, not the exact classifier reasoning for a particular episode. Oura’s explanation.
The local record now contains two concrete disagreements: Oura estimated 11m30s asleep around 22:20–22:46 on October 14 and 10m30s around 20:39–21:07 on October 15. I explicitly reported no dozing during either. The diary rejects them as actual naps; their original API classifications remain visible separately from main sleep. These are useful examples of disputed extra detections, not a measured error rate for every night or stage. October 15 record, October 16 record.
If an incorrect episode appears as a nap in the app, Oura says it can be deleted there to remove its effect on Sleep and Readiness Scores. The API’s short sleep episode is not automatically proof of how it appears in the app. The published score remains the returned score, without a hypothetical correction or an assumption about its exact contribution. Nap guidance.
What an accuracy percentage means
Binary sleep/wake classification and four-stage classification answer different questions. Sensitivity to sleep is the fraction of reference sleep detected as sleep. Wake detection, often called sleep specificity, concerns reference wake. Four-stage agreement requires the right label among wake, light, deep and REM.
Imagine a record with 90 sleep epochs and 10 awake epochs. A classifier that labels everything sleep has 90% binary accuracy, 100% sleep sensitivity and 0% wake detection. This mathematical example shows why a headline accuracy alone is insufficient.
For numerical quantities, examine mean bias and the spread of individual differences. Errors can cancel in a group average. Correlation can remain perfect when one device always adds a fixed offset. Limits of agreement describe the observed or modeled spread of paired differences under the tested conditions; they do not promise an exact error bound for every future user.
Brain states and sleep staging
PSG supplies the conventional reference using EEG, eye movement and muscle signals, with trained scoring. It is a reference method with its own scorer variability and protocol requirements. A ring’s optical and movement signals infer states indirectly.
The 2024 Gen3 / OSSA 2.0 study of 35 healthy adults found 92.0% binary agreement and 76.3% four-stage agreement. Wake sensitivity was 68.6%. Its duration results need the individual-agreement analysis, rather than only tests of group means. Oura funded it and the lead author disclosed Oura relationships. Full study.
The 96-person multi-night Gen3 study makes the distinction even clearer: its 90.6% REM accuracy was REM versus everything else; sensitivity to actual REM was 72.6% on the non-dominant hand. The hotel protocol used exploratory Gen3 devices modified for continuous PPG and manually recorded lights-out/on. Oura mainly funded it, with financial support disclosed by three authors. Its limits for total sleep, approximately −49 to +53 minutes on that hand, describe participant averages across available nights—not single-night Ring 5 precision. Svensson and colleagues.
The current Ring 5 product page’s 95% sleep-accuracy headline cites a 2021 paper and explicitly describes binary classification. The underlying research used a Gen2M research ring, 106 participants, 440 nights and cross-validated classifiers. That paper is not a Ring 5 deployment test. Its authors disclosed Oura employment/advisory relationships. Current product footnote, underlying paper.
A 2025 clinical-cohort study illustrates another boundary: Oura’s group-average total-sleep error was about 12 minutes, while individual errors ranged from −54 to +97.5 minutes. It analyzed 31 usable Oura nights from 45 patients, collected in 2022; exact ring generation and algorithm were not clearly specified. There was no specific study funding and no declared competing interests. Its numbers cannot calibrate our Ring 5. Herberger and colleagues.
An EEG headband can be a closer research reference, but needs its own validation. A 2020 Dreem study analyzed 25 simultaneous PSG nights with five expert scorers: automatic five-stage agreement was 83.5%, with expert agreement averaging 86.4%. Dreem supported the study and several authors were employees. These results belong to that tested system; they are not a validation of all headbands or a direct ranking against the different Oura cohorts. Arnal and colleagues.
Heart rate and beat intervals
ECG measures electrical beat timing; optical sensors measure peripheral pulse signals. A 2025 home comparison of earlier Oura generations used an H10 chest strap reference. Ring 4 nightly RHR limits of agreement were −3.75 to +1.87 bpm, with mean bias −0.94 bpm. Its cohort contained 13 participants, despite many nights. AFRL funded the study, with no declared competing interests. Dial and colleagues.
The same study’s Ring 4 HRV limits were −11.78 to +9.85 ms. A small average error or high concordance therefore does not make every individual night interchangeable with the reference. Its reference used artifact correction and excluded recordings with 5% or more corrected beats. Study methods and tables.
Quality filtering is part of the measurement. A Gen3 analysis of 114 participants found that stringent filtering improved HRV agreement while rejecting substantial data; shorter-window individual errors exceeded the apparent accuracy of nightly averages. Singapore public/philanthropic sources funded it, with no declared conflicts. Le and colleagues.
An electrical chest strap is a practical candidate for a personal HR/beat-interval comparison. A 2026 study tested the H10 raw intervals, 43 healthy young adults, against laboratory ECG during a short supine, breathing and standing protocol. It reported excellent agreement and no specific funding or conflicts. The result does not validate an arbitrary app’s artifact correction or every exercise/nighttime context. Blalock and colleagues.
Breathing and oxygen
Rate, oxygen saturation and detected desaturations need separate reference questions. Direct respiration measurement can use airflow or respiratory-effort channels. A nightly mean rate cannot measure every obstruction, while mean SpO₂ cannot expose the whole desaturation pattern.
Oura’s public API defines BDI from detected SpO₂ drops, as an integer in [0,100]. Its enterprise description discusses average detected disturbances per hour. The exact mapping, rounding and event rules for our public response remain unresolved. Neither establishes equivalence with clinically scored AHI. API documentation, manufacturer’s enterprise description.
When the question is suspected sleep-disordered breathing, clinically selected PSG or a technically adequate home study can assess it. AASM specifies appropriate patient selection and diagnostic pathways. A consumer breathing grade cannot substitute for these. This is a conditional reference choice, not a recommendation to order tests simply to generate more data. AASM diagnostic guideline.
Measuring biological phase
Clock time and internal circadian phase are separate. Dim-light melatonin onset (DLMO) assesses the rise of melatonin under controlled light and sampling conditions. It can resolve a phase question that a timing grade cannot.
A 2015 study tested monitored home collections against laboratory assessment in 35 healthy adults. Home onset averaged 9.6 minutes earlier; 92% of home assessments were not affected by identified light or sampling errors. NIH funded the work, with no financial conflicts. Compliance and assay procedures are central to interpreting it. Burgess and colleagues.
Finger temperature is neither a core-temperature trace nor a direct phase measurement. Room temperature needs separate measurement too. Temperature definitions. DLMO or specialized temperature instrumentation becomes worthwhile when resolving phase would change a consequential decision.
Measuring waking outcomes
A credible vigilance assessment offers an outcome alongside the night metrics. Published PVT-B evidence concerns a defined test, hardware and scoring, with limitations when the procedure changes. We have not selected a personal threshold or implemented a validated protocol. Functioning page.
Reaction performance, reported sleepiness, memory and training recovery should remain separately named. They need not agree every day. That disagreement can be informative rather than a reason to discard the less flattering measure.
Timing accuracy needs a measurable acceptance criterion. A published simulation proposed added latency bias within ±5 ms and variability ≤10 ms relative to its assumed reference recordings. These proposed apparatus goals do not certify a hardware/software combination, nor define our personal performance target. Calibration analysis. Its authors’ later response also encourages reporting measurement error and protocol when optimal margins cannot be achieved. Method discussion.
What comparing two wearables can establish
Repeated simultaneous observations can establish personal agreement, disagreement, missingness and observed variation. They can identify nights worth closer examination. Two devices can also share a bias. Agreement does not establish accuracy without an appropriate independent reference.
Between-night variation combines biological change and measurement effects. Its standard deviation should not automatically become a sensor-error estimate. Likewise, a stable number could reflect stable physiology or a measurement that misses change. Distinguishing these possibilities is a further reference question.
Outputs within one device can share inputs too. Oura’s staging algorithm uses cardiac and other peripheral signals. A change in estimated deep sleep alongside a change in the same ring’s HRV is therefore not two independent confirmations of a brain-state change. That is an inference from the measurement design, to be tested with a separate reference when it matters. Manufacturer’s signal-to-stage description.
Randomizing an intervention does not guarantee constant classification error across its conditions. A change that directly alters pulse, skin temperature or movement could affect both sleep physiology and the signals used to infer stages. This is a conditional measurement concern, not evidence that our Ring 5 has made a particular error. A relevant waking outcome provides a separate test of benefit; a consequential claim about brain stages may require simultaneous EEG or PSG.
For each comparison, retain the selected episode, timestamps and timezone, units, statistic, algorithm information, coverage and native values. Align periods before comparing. Report signed differences and their spread. Do not remove inconvenient nights, fill gaps with invented normality or average unlike quantities.
The decision also determines the right analysis. Agreement for a person’s average across several nights does not establish the same precision for a single night. Repeated nights belong to the same person; the analysis should account for that dependence. A 2024 comment and author reply debated this point for the multi-night Gen3 validation. The detailed reply has not been inspected here, so the critique remains a methodological question, not a settled verdict. Comment, reply.
Native resolution matters too. The API’s five-minute sleep-phase string cannot reproduce a paper’s thirty-second classification test. Its sleep-HR average and minimum use thirty-second samples and can differ from the app’s summaries of five-minute aggregates. A frontend or export that changes sampling is part of the measurement procedure. API reference.
The first night is an example: the devices’ sleep estimates differ by 14 minutes, and their nightly HRV summaries round to 61 ms. COROS’s predicted waking HR and Oura’s sleeping minimum should not be judged as competing readings of the same quantity. Cardiac definitions.
What would advance this investigation
First establish repeated ordinary nights, data completeness and a credible waking outcome. Next choose a consequential uncertainty and the lightest adequate reference. Direct EEG may earn its place for a stage-specific hypothesis; matched beat intervals for a cardiac hypothesis; DLMO for a circadian-phase hypothesis.
Improved measurement is itself progress when it changes what we can conclude. It should support a defined question and decision, rather than expand the instrument collection without a corresponding gain in understanding.