Smartwatches identify REM sleep by combining two main data streams: tiny movements detected by a motion sensor and changes in your heart rate captured by a light-based pulse sensor on the underside of the watch. Neither sensor can directly observe brain activity the way a clinical sleep study does, so the watch feeds both signals into a trained algorithm that estimates which sleep stage you’re in at any given moment. The result is the colorful sleep chart you see in your morning report.
The Two Core Sensors
Every smartwatch that tracks sleep stages relies on an accelerometer and a photoplethysmography (PPG) sensor working together. The accelerometer is a tiny chip that measures acceleration along three axes, picking up even small wrist twitches. It calculates an “activity count” each second by combining motion from all three directions and filtering out the constant pull of gravity. A three-second window is then classified as active or still based on whether any meaningful movement was detected.
The PPG sensor shines green light into your skin at a rapid rate, typically 25 times per second. As blood pulses through the capillaries in your wrist, the amount of light absorbed changes slightly with each heartbeat. The watch reads those fluctuations to derive your heart rate on a beat-to-beat basis, which also lets it calculate heart rate variability: the subtle differences in timing between consecutive beats.
What REM Looks Like to Your Watch
During REM sleep, your brain is highly active but your voluntary muscles are essentially paralyzed, a state called muscle atonia. For a healthy sleeper, that means the accelerometer picks up very little wrist movement during REM, similar to the stillness of deep sleep. Research comparing wrist-worn accelerometers to video and muscle-activity monitors found that in healthy people, movement levels during REM and non-REM stages showed no significant difference. So movement alone can’t reliably distinguish REM from other sleep stages.
Heart rate patterns are where the real signal lives. When you drop into non-REM sleep, your nervous system shifts toward a calm, rest-and-digest mode. Heart rate slows, and the beat-to-beat variability takes on a specific signature: the high-frequency component of that variability roughly doubles compared to wakefulness. When REM begins, the pattern flips. Your sympathetic nervous system (the “fight or flight” branch) ramps up, high-frequency variability drops back to waking levels, and overall heart rate variability increases. In a study of healthy subjects published by the American Heart Association, the ratio of low-frequency to high-frequency heart rate variability jumped from about 1.2 during non-REM sleep to about 3.0 during REM, nearly matching waking levels. Your watch is essentially looking for that nervous-system shift while you remain physically still.
How the Algorithm Puts It Together
Raw sensor data alone is noisy and ambiguous. To turn it into a sleep stage label, smartwatches run the data through machine learning models trained on thousands of nights of sleep that were simultaneously recorded with clinical polysomnography, the gold-standard test that uses brain electrodes, eye sensors, and muscle monitors. Google Research, for example, published a model that uses a 70-layer convolutional neural network fed with raw pulse and accelerometer signals, both sampled 25 times per second. The model was pretrained on over 1,600 clinical sleep recordings and outputs one of four labels for each 30-second window of the night: wake, light sleep, deep sleep, or REM.
The algorithm learns patterns that would be impossible to code by hand. It picks up on combinations of features: a particular heart rate trend appearing alongside near-zero movement, a characteristic rhythm in beat-to-beat intervals, or a subtle shift in pulse waveform shape. Each manufacturer trains its own proprietary model, which is why two watches worn on the same wrist can give you slightly different REM totals.
Extra Sensors Some Devices Use
A few devices layer in additional signals to refine their estimates. Skin temperature is one. During REM sleep, your body temporarily loses the ability to regulate its own temperature, a phenomenon called poikilothermia. Your skin temperature drifts toward room temperature, and devices like the Oura Ring, which sits snugly against the finger, can detect that shift. Breathing patterns also change during REM. Tidal volume (the amount of air in each breath) drops, and breathing becomes more irregular. Some wearables estimate respiratory rate from the pulse signal and use that information as another input.
Blood oxygen measurements, available on many newer watches, could theoretically help as well, since oxygen levels can fluctuate more during REM. However, most manufacturers haven’t publicly detailed how (or whether) SpO2 data feeds directly into their sleep staging models.
How Accurate Is REM Detection?
A validation study comparing three popular consumer devices to polysomnography found meaningful differences in REM sensitivity, which measures how often the device correctly identifies a REM epoch when clinical equipment confirms one. The Apple Watch Series 8 performed best, correctly flagging REM sleep 82.6% of the time. The Oura Ring came in at 76.0%, and the Fitbit Sense 2 at 67.3%. All three showed considerable person-to-person variation: the Fitbit’s standard deviation was 23.3 percentage points, meaning some users got much better results than others.
To put that in context, if you spent 90 minutes in true REM sleep, the Apple Watch would correctly label about 74 of those minutes as REM while misclassifying the remaining 16 minutes as another stage. That’s useful for spotting broad trends over weeks, like whether your REM sleep is gradually increasing after you cut back on alcohol. It’s less reliable for judging any single night down to the minute.
What Can Throw Off the Reading
Several factors can degrade accuracy. Skin tone matters: the green light used by PPG sensors is absorbed differently by melanin, and multiple studies have found that pulse readings are less accurate in people with darker skin. Since heart rate data is the primary input for distinguishing REM from other stages, reduced signal quality ripples through the entire classification. Tattoos on the wrist can cause similar issues by blocking or scattering light.
A loose-fitting watch introduces noise into both sensors. The accelerometer picks up the band sliding around as motion, and the PPG sensor loses consistent skin contact. Alcohol, certain medications, and sleep disorders can also alter the heart rate and movement patterns the algorithm expects, pushing its predictions further from reality. One notable example: people with REM sleep behavior disorder physically act out their dreams, producing bursts of movement during REM. In that population, wrist activity during REM is significantly higher than during non-REM sleep, which is the opposite of what the algorithm was trained to expect in healthy sleepers.
What Your Watch Can and Can’t Tell You
Your smartwatch gives you a reasonable approximation of your REM sleep, not a precise measurement. It’s reading indirect signals (pulse and movement) and using statistical models to infer what your brain is doing. A clinical sleep study measures brain waves directly, which is why it remains the standard for diagnosing sleep disorders. What your watch does well is track relative changes over time. If your weekly average REM percentage drops noticeably after a lifestyle change, that trend is likely real even if the nightly numbers aren’t perfectly accurate. Treat the data as a useful pattern tracker rather than a medical instrument, and it becomes a genuinely informative tool.