Why Sleep/Wake Is Easier Than Sleep Stages
Once a smart ring’s sensing chain is visible, the next question is whether its labels agree with polysomnography (PSG). The useful comparison is not a single brand score but the target the ring is trying to classify: broad sleep versus wake, or the finer architecture of light, deep, and rapid eye movement (REM) sleep. Reading the evidence at that level shows what a ring can reasonably support and where a precise-looking timeline becomes an uncertain inference.
A strong binary result still has a weak side
A two-class decision asks the ring to divide a period into wake or sleep. That is a simpler target than distinguishing several neighboring physiological states. It can use broad changes in movement and cardiovascular signals, even though those clues remain indirect. The question is not whether the ring has learned anything about sleep. It has. The question is how much meaning to assign to the particular agreement number.
In a meta-analysis of seven studies comparing finger-worn devices with PSG, pooled accuracy for wake/sleep classification was 0.87, with a 95% confidence interval from 0.86 to 0.89. In everyday language, the studies’ epoch-level classifications were correct about 87% of the time on average under their study conditions. The confidence interval describes uncertainty around that pooled estimate; it does not describe the range of nights a particular person should expect, and it does not turn the result into a product-wide guarantee.
The individual results make that qualification necessary. One Oura Ring configuration reported accuracy of 0.93, while a Circul Ring study reported 0.66. Those numbers are not necessarily contradictory. They may reflect different devices, algorithms, participant groups, settings, and validation procedures. A pooled result summarizes the available studies; it does not erase the conditions that produced each one.
There is another reason not to read 0.87 as a promise about a personal night. A confidence interval around a pooled estimate describes uncertainty in the synthesis of studies; it is not a prediction interval for one user’s next sleep period. Nor does it tell us whether the errors were scattered harmlessly across short intervals or concentrated around awakenings and transitions. Two devices could have the same accuracy while producing different kinds of mistakes, and those mistakes would matter differently to someone trying to understand fragmented sleep.
A nightly total can also conceal timing errors. If one quiet-awake period is labeled sleep and one sleeping period is labeled wake, the total may look roughly plausible even though the timeline is wrong. That may be acceptable for a broad pattern question and unacceptable for a question about repeated awakenings. The metric does not decide the use case; it only describes agreement according to the study’s chosen unit of analysis.
Nor does accuracy answer every relevant question. It is the proportion of all classifications—correct and incorrect—that were correct. Sensitivity asks how well a test identifies the class treated as the condition, while specificity asks how well it excludes that class when it is absent. Balanced accuracy averages the two class-specific performance measures. A system can have a respectable overall accuracy while making many errors in one direction, especially when one class occupies more time or is easier to recognize than the other.
The review reported two-stage sensitivity ranging from 0.74 to 0.97, specificity from 0.35 to 0.98, and balanced accuracy from 0.55 to 0.90. The wide ranges are not a footnote to the pooled 0.87; they explain why the pooled number needs interpretation. A ring may be good at recognizing obvious wakefulness while treating quiet wake as sleep, or it may mark brief movement as wake when the person is still asleep. Which error matters depends on the use: estimating a broad nightly pattern is different from deciding whether every awakening was detected.
The quiet-wake problem follows directly from the ring’s inputs. If movement is a major clue, wakefulness without movement resembles sleep. A person can be mentally alert while lying still, reading, worrying, or waiting to fall asleep. The accelerometer has not malfunctioned in that situation. It has correctly recorded little movement, but little movement does not have a unique meaning. The algorithm must infer the state from whatever other signals are available, and the inference can still be wrong.
The reverse error is possible too. A turn, a hand movement, or a temporary pulse change may look like wakefulness even when the person remains asleep. This is one reason a sleep timeline should not be read as a literal replay of consciousness. It is a sequence of model decisions about time windows. The more a user treats every short segment as an exact event, the more the unavoidable ambiguity matters.
The evidence is therefore encouraging for a limited claim: finger-worn devices can provide a useful estimate of broad sleep and wake patterns, especially when the goal is repeated observation rather than clinical diagnosis. It is weaker for the stronger claim that a ring knows exactly when a person fell asleep, exactly how long they were awake, or exactly what a proprietary morning score means. Epoch-level agreement with PSG is evidence about a classification task, not proof that the app has directly measured a person’s entire night.
Sleep stages expose the signal the ring does not directly have
Four-stage classification asks for a much finer distinction. Clinical sleep scoring separates non-REM stages such as N1, N2, and N3 from REM sleep using characteristic patterns of brain activity, alongside other physiological observations. A typical smart ring does not directly record those brain-wave patterns. It observes peripheral consequences—movement, pulse dynamics, oxygenation, temperature, and related features—and uses them to infer which stage is more likely.
That difference in target and signal explains the performance gap better than the label “artificial intelligence” does. A model can be sophisticated and still be limited by the information available to it. If two stages produce overlapping peripheral patterns, the algorithm must draw a boundary using clues that are correlated with the defining physiology rather than observing the defining physiology itself.
The pooled results reflect that difficulty. Across five studies, accuracy was 0.65 for light sleep, 0.81 for deep sleep, and 0.74 for REM sleep. The confidence intervals were 0.56–0.74, 0.77–0.85, and 0.68–0.79, respectively. Reported sensitivities ranged from 0.45 to 0.82 for light sleep, 0.17 to 0.89 for deep sleep, and 0.14 to 0.89 for REM sleep. Cohen’s kappa, a measure of agreement that adjusts for agreement expected by chance, ranged from 0.06 to 0.70 where it was reported.
The stage-specific numbers should not be turned into a simple ranking of sleep stages. Deep sleep’s pooled accuracy being higher than light sleep’s does not mean the ring has a direct deep-sleep sensor. It may mean that the peripheral patterns, model, participant group, and study procedures made that category easier to separate. Similarly, a low result for one stage does not prove that every device is equally weak. It shows that the task remains uneven and that the output must be interpreted as a classification estimate.
Individual examples show both progress and limits. A personalized Happy Ring algorithm reported 0.81 accuracy for light sleep, 0.95 for deep sleep, and 0.92 for REM in one study. By contrast, an Oura Ring Generation 2 study reported 0.52 for light sleep and 0.51 for REM, while an earlier Oura generation reported 0.51 for deep sleep. The lesson is not that one named product permanently succeeds and another permanently fails. The lesson is that algorithm version, personalization, sample, and study design can move the number substantially.
A thirty-second epoch can also change the reported result. In a subgroup analysis, studies using 30-second epochs had pooled accuracy of 0.88 for two-stage sleep and 0.66 for light sleep. But the corresponding I² values—94.4% and 99.6%—were very high. I² is a measure of how much variation remains between study results beyond what sampling error would ordinarily explain. High heterogeneity means that sharing an epoch length did not make the studies equivalent. It would be a mistake to read the subgroup estimate as proof that choosing 30 seconds causes a particular improvement for an individual ring user. Epoch length changes the unit being compared, but it does not settle how the algorithm handles transitions, ambiguous intervals, or the signals available in each participant.
The reference itself also shapes the result. An epoch-by-epoch comparison asks whether the ring and PSG assigned the same category to the same interval. That is stricter in one sense than comparing a total sleep-time estimate, but it is not the only measure of usefulness. A device might approximate a nightly total while missing the exact timing of brief awakenings, or show reasonable agreement on broad states while struggling with transitions. Those are different claims and should not be substituted for one another.
Participant characteristics matter for the same reason. The review found higher two-stage accuracy in predominantly healthy participants, 0.89, than in unhealthy participants, 0.78. Participants aged 18–50 generally showed higher accuracy across most sleep stages than participants over 50, and one comparison found little effect from hand dominance. These results do not say that age alone determines a ring’s performance. They show that a validation result belongs to a population and that physiology, comorbidity, and study composition can alter how well the model’s learned patterns transfer.
The evidence base itself also limits confidence. Of the 28 studies in the finger-worn review, only one was judged to have low risk of bias overall; 27 had some risk. Seventy-one percent were judged high risk in the flow-and-timing domain, often because participants were excluded for reasons such as insufficient sleep duration or device malfunction. Participant selection was frequently unclear. These are not reasons to discard every result. They are reasons not to mistake a collection of heterogeneous validation studies for a standardized certification test. Excluding nights with insufficient sleep or device problems can make the remaining sample easier to classify than ordinary use. The review therefore retains studies with identified risks of bias, but its numerical conclusions have to travel with that uncertainty. A result can be directionally useful without being precise enough to settle a clinical or personal question.
The population gap is just as important as the statistical gap. Sleep-stage evidence was concentrated in healthy adults, while apnea evidence was concentrated in people already suspected of having the disorder. Those are sensible groups for early validation, but they are not the whole market of ring users. A healthy younger participant, an older adult with another condition, and a child may present different signal patterns and different consequences when a label is wrong. “Validated against PSG” is therefore incomplete unless the population and task are named.
The same caution applies when a manufacturer releases a better-sounding version. A change in hardware, firmware, feature selection, or proprietary algorithm can alter the output. An improvement reported for one version is evidence about that version and its validation conditions. It is not automatically evidence about a later version, an earlier version, or a different ring with a similar sensor list. Better engineering can raise agreement without resolving the underlying fact that the ring is inferring neural sleep stages from peripheral signals.
The evidence can be read with a practical set of questions. What was the target—wake/sleep or a specific stage? What was the reference—PSG, actigraphy, or something else? Were the reported figures accuracy, sensitivity, specificity, or chance-adjusted agreement? Which device and algorithm version were tested? How long was each epoch? Who participated, who was excluded, and where did the recording occur? A number becomes more informative when these conditions travel with it. Without them, “87% accurate” sounds like a stable property when it is really a conditional observation.
The diagram is not saying that the binary task is easy or that every stage estimate is useless. It shows why the tasks have different information requirements. A ring can collect several peripheral clues and still lack the signal most directly tied to the stage label. That is a structural limit, not merely a temporary software defect.
This distinction also prevents an apparent contradiction in consumer sleep technology. A later algorithm may outperform an earlier one, multisensor fusion may improve a validation result, and a ring may still be unsuitable for detailed sleep-architecture assessment. Improvement means the model is extracting more value from the available signals. It does not mean that the available signals have become the same as PSG’s brain-wave record.
A smart ring’s performance is therefore not a general competence score. It is a relationship between a target, a signal set, a model, a population, and a reference standard. Sleep/wake tracking asks for a broad separation and can be useful at that level. Stage scoring asks for finer distinctions and remains less settled. Breathing-event detection will test the same principle differently, because it depends on oxygenation and respiratory patterns and on the clinical threshold used to define a meaningful flag.