I hold a PhD in neuroscience and have spent 13 years in sleep research. Before joining Oura, I was a postdoctoral researcher at UC Berkeley where I built an open-source algorithm that scores sleep stages from polysomnography. It was published in a top-tier journal and is now among the most widely used and cited automated sleep staging tools in the field.
Since 2020, I have been the lead algorithm scientist for sleep tracking at Oura. That work included building the sleep staging algorithm the ring uses today, trained on one of the largest sleep wearable datasets in the field, deliberately spanning different ages, skin tones, health statuses and sleep disorders.
A proposed class action filed against Oura last week alleges that our models are ”doing guesswork” and that our sleep staging has ”a coin flip’s chance of being correct.” These claims are made on the basis of a 2025 study, published in Scientific Reports, that reported a 53% agreement for four-stage classification with the Oura Ring Gen3. That number comes from a single study, which also has several methodological flaws that my colleagues and I addressed in 2025 in a peer-reviewed letter to the editor in Sleep Advances,”Toward better evaluation of consumer sleep technologies: a call for rigor, context, and collaboration” (more on these later). Scientific progress depends on open, rigorous debate. For generations, scientists have tested, challenged, and refined one another’s work through the peer-reviewed literature. That process—not headlines—is how evidence is evaluated and how the field moves forward.
But let’s go back to arithmetic. A coin flip has two possible outcomes. Wearables like Oura typically score sleep into four stages, assigning each moment of the night to wake, light, deep, or REM. If a model simply guessed at random among these four stages, its expected accuracy would be 25%. In other words, the chance level for four-stage classification is 25%, not the 50% that the coin flip analogy implies. For context, recent studies, both Oura-funded and independent, report four-stage agreement with clinical polysomnography in the 70 to 79% range in healthy adults, and sleep versus wake agreement of ≥90%. Not one study, ours or anyone else’s, reports a number anywhere close to 25%. We have summarized these validation studies, along with a Q&A on how sleep staging works, in this blog article, “Standing Behind Our Science: How Oura Measures Sleep and Validates Accuracy.”
Setting aside the basic math error, there are several issues with the Scientific Reports study, as I mentioned earlier:
- Improper use conditions: The study did not follow Oura’s guidance on ring fit and finger placement, using just two ring sizes (8 and 12) across all study participants. This likely explains why nearly a third of the participants produced no usable Oura data at all in this study (14 out of 45 participants), which is a much higher failure rate than in any other study that has evaluated Oura against PSG. The study reported that “dropouts occurred due to […] problems with ring fit, such as rings that fell off or were taken off prematurely due to poor fit.” It goes without saying that ring fit is critical for signal quality, and that a ring that falls off the finger is unlikely to produce reliable results. Yet, this data loss is interpreted as an inherent limitation of ring trackers rather than as a consequence of suboptimal experimental conditions.
- Non-standard evaluation methods: The study did not follow the field’s established methods for evaluating wearable sleep trackers (e.g. comparing 5-minute Oura hypnograms versus 30-second PSG hypnograms, thus artificially lowering epoch-by-epoch agreement; or not padding missing device epochs to PSG bedtimes—see footnotes for details).
- Outdated algorithm and device version: The paper was submitted more than two years after the data were collected. In that interval, Oura had made a major upgrade to its sleep staging algorithm and launched a new generation of the ring with a redesigned sensor platform. Neither update is mentioned anywhere in the paper (see footnotes).
- Overinterpretation of the results: None of the limitations above are weighed in the paper’s discussion, yet the authors draw categorical conclusions about the value of ring trackers as a whole. Conclusions of that scope require replication and should not rest on a single study of 31 usable nights.
Our models are not “guesswork.” We have spent years collecting thousands of nights of simultaneous clinical polysomnography and ring data to train and validate our algorithms. The model learns the mapping between those polysomnography-derived expert labels and the peripheral signals a ring can capture, including heart rate, heart rate variability, breathing rate, temperature and motion. All of these vary across sleep stages in distinct and reproducible ways, as decades of research on autonomic activity and brain-body coupling during sleep have documented.
In 2021, Oura became one of the first wearable companies to transparently publish the inner workings of our algorithm in a peer-reviewed journal: the model, the training data, and how much each sensor contributes to the accuracy. That paper has since been cited more than 250 times in peer-reviewed scientific articles.
The accuracy of wearable sleep trackers is measured against human expert scoring of polysomnography — the current gold standard for measuring sleep. Early in my career I manually scored dozens of overnight polysomnography recordings. You work through the night in 30-second windows, called epochs, assigning a stage to each one, which comes to around a thousand decisions per night. Anyone who has done it remembers how difficult and ambiguous the task can be, particularly around stage transitions or in individuals with fragmented or disordered sleep. Sleep is a continuous process, and a single 30-second epoch can contain the signatures of two different stages, leaving the scorer to decide which one to score.
Give two trained experts the same night of polysomnography and they will disagree on roughly one epoch in six: agreement between scorers is around 83% in healthy adults, and lower in people with sleep disorders. So the target for any sleep staging algorithm was never 100%, but rather the level of agreement that two trained humans reach with one another.
An imperfect reference standard is not an excuse for an imperfect algorithm. It means two things have to keep getting better: the algorithm itself, and the way people evaluate it. We have plenty of work ahead of us on both counts. Performance in older adults and in clinical populations is one area my team and I are actively working to improve, and we invest continuously in research and development to push accuracy higher and to make it consistent across populations and sleep conditions.
Oura has consistently engaged with the scientific community, supported independent research, and helped develop better standards for how these devices are evaluated. Our letter in Sleep Advances sets out what a rigorous, technologically current, and contextually grounded evaluation framework should look like, from version reporting to population diversity. It is a standard we intend to be held to as much as anyone else.
Footnotes:
Non-standard evaluation methods
Five-minute device output was compared against 30-second reference scoring. Oura’s public API reports sleep stages in five-minute blocks, and the authors assigned each five-minute label to all ten underlying 30-second epochs scored by polysomnography. This imposes a strong ceiling on performance as any short stage transition (e.g. a 1-min awakening) is smoothed out and counted as a misclassification. The native 30-second Oura output is available to researchers and has been used in many previous evaluations.
Missing device epochs were discarded rather than scored as wake. Where ring data was absent at the beginning or end of a recording, those epochs were excluded from the agreement analysis. The conventional practice is to pad to the polysomnography window and label them wake, since by definition an epoch outside the device’s detected sleep period is wake. Discarding them removes precisely the epochs where a device is most likely to be correct about wake, further penalizing wake sensitivity, the metric the authors then highlight as weakest.
Other non-optimal methodological choices include:
- Lack of consensus scoring: each night was scored by a single technician with no consensus procedure. Human scorers agree with each other roughly 83% of the time on five-stage classification. Interscorer variability is an intrinsic limitation of the reference measure, and should be recognized when interpreting device performance.
- Lack of time-shift correction: even minor clock offsets between the device and the reference PSG system can substantially reduce the epoch-by-epoch agreement. Standard procedures for evaluating wearable performance include a time-shift estimation and correction step, which was not performed in this study.
Outdated algorithm and device version
The devices section of the paper identifies the Oura ring by size and generation, but does not report the firmware, app, or algorithm versions. The data for this study were collected in 2022, a few months before Oura started rolling out a major upgrade to its sleep staging algorithm. This upgrade delivered a step change in accuracy, and it was documented at length across our blog, member and partner communications, and peer-reviewed validation studies. The manuscript was submitted in January 2025, nineteen months after the algorithm release, with no mention of it anywhere in the paper. The Oura Ring 4 had also launched in October 2024, three months before submission, and is similarly absent from the discussion. The results should have been contextualized as reflecting a hardware and software generation that was no longer in use by the time of writing.