How Do Wearables Detect Sleep Stages? Sleep Stage Detection Explained
Wearables detect sleep stages by dividing the night into short time windows and analysing patterns in movement, heart rate, heart-rate variability and other available signals such as temperature or blood oxygen. Algorithms trained against polysomnography-labelled sleep use these indirect clues to estimate whether each window is wake, light sleep, deep sleep or REM. Because most wrist- and finger-worn devices do not measure brain waves, eye movements or muscle tone, their sleep-stage results are estimates, and are generally less reliable than broad sleep-versus-wake detection.
A sleep graph can make the night look wonderfully orderly: light sleep at 11:48 pm, deep sleep at 12:22 am and REM at 2:06 am. But your ring or watch is not directly watching your brain move through those stages. It is solving a classification problem from the signals available at the wrist or finger.
Our previous article, How Do Wearables Know When You’re Asleep?, explained broad sleep detection. This article goes one level deeper: how those devices turn movement and cardiovascular patterns into a stage-by-stage estimate of your night.
Sleep Stages Explained: Wake, Light, Deep and REM
Clinical sleep scoring uses five labels: wake, N1, N2, N3 and REM. Consumer apps often simplify them to awake, light, deep and REM, combining N1 and N2 into “light sleep.”
N1: The Transition into Sleep
N1 is a brief, light transitional stage. Brain activity begins moving away from relaxed wakefulness, slow eye movements may appear and muscle tone starts to fall. It is easy to wake from N1, and people may not always feel that they were asleep.
N2: Stable Light Sleep
N2 usually forms the largest share of adult sleep. In a sleep laboratory, it is identified by distinctive EEG events called sleep spindles and K-complexes. A typical consumer wearable cannot see either event directly, which is one reason it usually groups N1 and N2 together.
N3: Deep or Slow-wave Sleep
N3 is the deepest NREM stage. It is defined clinically by high-amplitude, low-frequency brain activity. Movement is generally low and parasympathetic (or, rest and recover) activity tends to dominate. Deep sleep is usually concentrated earlier in the night.
REM: Active Brain, Quiet Body
REM sleep combines wake-like brain activity with rapid eye movements and marked loss of skeletal-muscle tone. Heart rate and breathing can become more variable, and REM periods generally lengthen toward morning. The NIH overview of sleep physiology describes how adults typically move through four or five NREM–REM cycles, with later cycles lasting roughly 90–120 minutes.¹
How are Sleep Stages Measured Clinically?
The reference method is polysomnography (PSG). It combines EEG for brain activity, EOG for eye movements and EMG for muscle tone, usually alongside heart rhythm, airflow, respiratory effort, oxygen saturation and body position. Trained scorers review these signals in 30-second windows, called epochs, and assign each epoch a stage using standard criteria.
That matters because sleep stages are defined primarily by signals consumer wearables do not normally record. N2 depends on EEG spindles and K-complexes; N3 depends on slow-wave brain activity; REM depends partly on eye movements and muscle atonia. A wrist or finger device must infer these states from their effects elsewhere in the body.
Even PSG labels are not perfectly unambiguous. In the AASM inter-scorer reliability programme, overall agreement between human scorers was 82.6%; agreement was lower for N1 and N3 than for several other stages.² Stage boundaries are physiological transitions, while scoring converts them into discrete labels. A wearable model is therefore learning from a useful but not infallible reference.
How do Wearables Detect Sleep Stages?
Most systems follow the same broad pipeline, although the implementation differs by manufacturer.
Sensors collect overnight signals
An accelerometer records movement. A PPG sensor uses light to capture a pulse waveform, from which heart rate and beat-to-beat intervals can be estimated. Some devices add skin temperature, SpO₂, respiratory features or other signals. Sensor placement and mechanical fit further influences the quality of everything that follows.
The raw data is cleaned
Motion can distort optical measurements. Loose contact, ambient light, device rotation and missing samples can also create noise. Signal-processing software filters artefacts, assesses signal quality and may exclude sections it cannot trust. A sophisticated algorithm cannot recover information that was never captured clearly.
The night is divided into epochs
Wearable data is commonly aligned to 30-second epochs so predictions can be compared with PSG scoring. The system may still examine minutes of context around each epoch. Think of a word in a sentence: its meaning is clearer when you read the words before and after it.
Useful features are extracted
Instead of treating millions of sensor samples as unrelated numbers, the system derives features such as movement intensity, heart rate, interbeat intervals, HRV, pulse-wave characteristics, temperature change, estimated breathing rate and time-of-night context.
Each clue contributes something different. Movement is strong for separating active wake from sleep but weak for distinguishing NREM from REM. Cardiac features become more important for multiple stages. Research on HRV has found a general shift toward greater parasympathetic influence across NREM sleep and more variable sympathetic activity during REM and wake, although individual patterns overlap and can change around arousals.
A classifier estimates the stage
A model trained on wearable recordings collected alongside PSG learns which feature combinations tend to occur with each reference stage. It may be a statistical classifier, gradient-boosting model, neural network or another machine-learning system. Its output is usually a probability for each class; not absolute certainty. For example: 12% wake, 55% light, 10% deep and 23% REM.
Temporal rules refine the sequence
Sleep has structure. An isolated deep-sleep prediction between wake epochs may be less plausible than a short continuation of light sleep. Models can use neighbouring epochs, transition probabilities, circadian timing and post-processing rules to reduce erratic stage switching. This can make the graph look more physiological, but excessive smoothing may also hide brief awakenings.
The app turns stages into metrics
Finally, the estimated labels become a hypnogram, stage durations and percentages. Some apps then use these estimates within a sleep score. That score is a further layer of interpretation, not a directly measured physiological quantity.
From Sensor Signals to Sleep Stages: A Real Overnight Example
The easiest way to understand sleep-stage detection is to compare wearable measurements and estimated sleep stages on the same timeline. In this overnight example, the heart-rate trace is synchronized with a hypnogram showing when the algorithm classified the wearer as awake or in light, deep or REM sleep.
The beginning of the recording shows a generally higher and more variable heart rate. Later, heart rate settles into a lower range, but the estimated stages continue to change. Light, deep and REM sleep do not each produce a distinct heart-rate band. Similar values can appear across several stages, while short-lived increases and decreases can occur without a corresponding stage transition.
This overlap is important. A wearable cannot identify a sleep stage by applying a simple rule such as “a heart rate below this value means deep sleep.” Instead, the algorithm examines features such as heart-rate level, short-term variation and the direction of change within consecutive epochs. Depending on the device, it may combine these with HRV, movement, respiratory patterns, temperature and signal-quality information.
The model then considers how these features occur together and how the current epoch relates to those before and after it. This broader context helps it choose the most likely stage and avoid physiologically unlikely sequences. The final hypnogram is therefore the result of interpreting several overlapping patterns; not reading a sleep stage directly from one vital-sign graph.
How Accurate is Wearable Sleep-Stage Detection?
A 2024 review of 35 studies and 62 wearable setups reported average accuracy of 87.2% for basic sleep/wake classification. Average accuracy fell to 69.7% for three classes: wake, NREM and REM, and 65.2% for four classes (wake, light, deep and REM.) Reported four-stage accuracies in the reviewed studies ranged from 69% to 79% where accuracy was available.³
A computationally efficient four-class model using PPG-derived cardiac activity and movement achieved 77.8% epoch-by-epoch accuracy and a median Cohen’s kappa of 0.638 against PSG on its hold-out data.⁴ Kappa estimates agreement beyond chance; a value around 0.64 indicates useful but imperfect correspondence, not clinical interchangeability.
Consumer-device results also vary by stage and system. A multicentre comparison of 11 sleep trackers analysed 349,114 epochs from 75 participants and found overall macro F1 scores ranging from 0.26 to 0.69.⁵ Different systems performed best for different stages, while errors in sleep efficiency and sleep latency also varied widely. In other words, “sleep-stage accuracy” cannot be reduced to one universal percentage.
Why Sleep-Stage Results Vary Between Wearables
Two devices can observe the same night and produce different graphs without either one being obviously broken. Differences may come from:
Sensor stack: Movement alone provides less stage information than movement plus good-quality PPG and contextual sensors.
Body location and fit: Finger, wrist and chest signals have different strengths and artefacts.
Sampling and power strategy: Battery limits affect how frequently and for how long signals are captured.
Stage definitions: One app may report N1 and N2 as light sleep; another may use a different internal mapping.
Training population: Age, health, skin characteristics, sleep disorders and medication use can affect generalisation.
Model and features: Algorithms weigh movement, cardiac, circadian and contextual inputs differently.
Temporal smoothing: One system may preserve rapid transitions while another produces a cleaner, smoother hypnogram.
Firmware updates: The same hardware can produce different results after an algorithm changes.
For product teams, this means testing the wearable with its intended users and in realistic conditions; not assuming that an accuracy result from a different device, population or dataset will apply to their product.
How much Light, Deep and REM Sleep is enough?
There is no clinically prescribed nightly quota for each stage that every adult must hit. Stage distribution varies with age, total sleep time, recent sleep, circadian timing, stress, alcohol, medication, illness and normal night-to-night biology.
As a broad reference, adult sleep is often described as approximately 5% N1, 50% N2, 20% N3 and 20–25% REM. Some reviews report wider ranges of 5–10% N1, 45–55% N2, 15–25% N3 and 20–25% REM.⁶ These are descriptive population-level estimates, and not nightly targets. Proportions vary with age, health, medication and normal night-to-night variation. The NIH summary of age-related sleep architecture further notes that slow-wave sleep declines markedly from young adulthood into midlife, while lighter sleep and fragmentation tend to increase with age.⁷
Avoid judging sleep by one stage estimate from a single night. Total sleep, timing, continuity, daytime alertness and multi-night trends usually provide a more useful picture. Persistent symptoms such as severe daytime sleepiness, gasping, loud snoring or chronic insomnia, deserve clinical assessment regardless of what the graph says. The AASM position on consumer sleep technology states that consumer devices should not be used to diagnose or treat sleep disorders without appropriate clinical evaluation.⁸
Conclusion
Wearable sleep staging begins with indirect signals and ends with an estimate. Its reliability depends on the entire chain: skin contact, sensing physics, firmware, artefact handling, features, model design, temporal context, validation and the way results are communicated.
The best systems do not pretend that one pulse pattern “is” deep sleep. They combine multiple clues, recognise how those clues evolve through the night and remain honest about uncertainty. That is what turns a colourful sleep graph into a useful physiological insight.
Build sleep insights that can be trusted
Sensio helps wearable teams move from sensor selection and signal processing to algorithms, prototypes, applications and validation. If you are developing a smart ring, band, patch or custom sleep-monitoring product, talk to us about building the complete system behind the insight.
References:
Brinkman JE, Reddy V, Sharma S. Physiology of Sleep. [Updated 2023 Apr 3]. In: StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; 2026 Jan-. Available from: https://www.ncbi.nlm.nih.gov/books/NBK482512/
Rosenberg RS, Van Hout S. The American Academy of Sleep Medicine inter-scorer reliability program: sleep stage scoring. J Clin Sleep Med. 2013 Jan 15;9(1):81-7. doi: 10.5664/jcsm.2350. PMID: 23319910; PMCID: PMC3525994.
Birrer, V., Elgendi, M., Lambercy, O. et al. Evaluating reliability in wearable devices for sleep staging. npj Digit. Med. 7, 74 (2024). https://doi.org/10.1038/s41746-024-01016-9
Fonseca, P., Ross, M., Cerny, A. et al. A computationally efficient algorithm for wearable sleep staging in clinical populations. Sci Rep 13, 9182 (2023). https://doi.org/10.1038/s41598-023-36444-2
Lee T, Cho Y, Cha K, Jung J, Cho J, Kim H, Kim D, Hong J, Lee D, Keum M, Kushida C, Yoon I, Kim J. Accuracy of 11 Wearable, Nearable, and Airable Consumer Sleep Trackers: Prospective Multicenter Validation Study. MIR Mhealth Uhealth 2023;11:e50983. URL: https://mhealth.jmir.org/2023/1/e50983. DOI: 10.2196/50983
Su X, Wang DX. Improve postoperative sleep: what can we do? Curr Opin Anaesthesiol. 2018 Feb;31(1):83-88. doi: 10.1097/ACO.0000000000000538. PMID: 29120927; PMCID: PMC5768217.
Nayak CS, Sankari A, Anilkumar AC. EEG Normal Sleep. [Updated 2026 May 20]. In: StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; 2026 Jan-. [Table], Table 2. Sleep Architecture Changes by Age. Available from: https://www.ncbi.nlm.nih.gov/books/NBK537023/table/article-20917.table1/
Consumer Sleep Technology: AASM position statement. American Academy of Sleep Medicine – Association for Sleep Clinicians and Researchers. (2019, October 8). https://aasm.org/advocacy/position-statements/consumer-sleep-technology/
Comments