Watch or chest strap? What PPG does to HRV numbers

The diagram below lines up the two signals for the same heartbeats: a chest strap’s ECG ECG ECG A recording of the heart's electrical activity, taken with electrodes on the skin. ECG-based RR intervals are the reference standard for measuring HRV. More about ECG above, a watch’s PPG PPG PPG An optical method that shines light into the skin and detects the pulse of blood with each heartbeat. Most watches, rings and phone-camera apps use it to estimate HRV. More about PPG below.

1. Pulse transit time: the pulse wave reaches the wrist a fraction of a second after the R peak.
2. The transit time changes with breathing and blood pressure, so a PP interval is not an exact copy of its RR interval.
3. Timing spread: a rounded wave can be timed at its foot, its steepest rise or its peak, and each gives a slightly different beat time.
4. Low sampling rate: at 25 samples a second the sensor sees one point every 40 ms, so the peak time is rounded.

Two traces, one above the other, for the same heartbeats. The top trace is an ECG with a tall, sharp R peak for each beat; brackets between the peaks are labelled RR interval. The bottom trace is a PPG with a smooth, rounded pulse wave for each beat. Each pulse starts a short time after its R peak, and that delay, the pulse transit time, is a little different for every beat. On one pulse a shaded band covers the foot, the steepest rise and the peak, the points a detector might use as the beat time. On another pulse, dots mark the samples a slow sensor takes, 40 ms apart, none of them exactly on the peak. Brackets between the pulse peaks are labelled PP interval.

People often ask whether a watch or ring measures HRV as well as a chest strap. The short answer is close at rest, and worse once you move. This page explains why, step by step: how a wearable turns pulses of light into a list of beat intervals, where timing error enters at each step, and which HRV numbers that error changes most. It is the deep dive behind the comparison on how HRV is measured. For the pulse wave itself and the other things a PPG sensor measures, start with PPG: how your watch reads your pulse.

Jump to the interactive explorer

Two signals from one heartbeat

A chest strap is a small ECG. Two electrodes pick up the electrical wave that spreads through the heart muscle just before it contracts. On the trace, each beat makes a tall, narrow spike, the R peak, that rises and falls in a few hundredths of a second. Its top is easy to place in time, so the gap from one R peak to the next, the RR interval RR interval RR interval The time from one heartbeat to the next, measured in milliseconds between the R peaks of an ECG. HRV is calculated from a series of these intervals. More about RR interval , is a precise record of when the heart’s pacemaker pacemaker Sinus node The heart's natural pacemaker, a small patch of cells that starts every normal heartbeat. Left alone it fires about 100 times a minute; the vagus nerve holds it lower at rest. More about Sinus node fired (Task Force, 1996).

A watch or ring uses PPG, an optical sensor. A light shines into the skin and a photodiode measures how much comes back. When a beat pushes blood into the small vessels under the sensor, a little more light is absorbed, and the sensor sees one smooth hump per beat. The gap between humps is the pulse-to-pulse (PP) interval, and the variation in those gaps is pulse rate variability pulse rate variability Pulse rate variability Variation in the time between pulse beats, measured optically by watches, rings and phone cameras. It closely tracks HRV at rest but drifts from it during movement. More about Pulse rate variability , PRV.

Three differences matter for timing:

• The pulse arrives later. The pressure wave has to leave the heart and travel down the arteries. In 32 reclining adults, the time from the R peak to the foot of the pulse was 126 ms at the ear and 269 ms at the finger (Block and colleagues, 2020). The wrist sits between those, so a watch probably sees each pulse roughly 150 to 300 ms after the R peak, depending on the person and the point timed on the pulse. That range is our estimate, not a measured value.
• The delay changes. Transit time shortens when blood pressure rises and lengthens when it falls (Mukkamala and colleagues, 2015), and it swings with breathing. During paced breathing in 15 healthy people, the largest change in transit time ranged from 7 to 23 ms between people (Drinnan and colleagues, 2001). Those milliseconds land in the PP intervals, but the heart never produced them.
• The pulse is round. The R peak climbs and falls in a few tens of milliseconds. The pulse wave takes about a tenth of a second just to rise, and its top is broad and flat. Placing a broad top in time is harder than placing a narrow spike.

Step by step: from light to an HRV number

Every wearable runs some version of this chain. Makers do not publish the details, so take the steps as the general method rather than any one product’s.

The diagram below shows the seven steps, numbered as in the headings that follow, with what can go wrong at each.

1. 1 Light in, raw signal out An LED lights the skin and a photodiode reads the light that returns. Accuracy lost here: A loose strap, daylight, cold skin or dark tattoo ink can shrink the ripple until it is barely above the sensor’s own noise.
2. 2 Sampling The sensor takes a reading at fixed ticks, the sampling rate. Accuracy lost here: A pulse can only be placed on a tick, so each pulse time is rounded.
3. 3 Filtering Fast noise is smoothed and slow drift removed. Accuracy lost here: Smoothing blunts each pulse; a shift that depends on its shape moves the beat time.
4. 4 Finding the same point on every pulse The detector times one landmark: the foot, the steepest rise or the peak. Accuracy lost here: Noise moves a broad, flat peak more than a steep upstroke.
5. 5 The interval series Each pulse time is subtracted from the next: one number per beat. Accuracy lost here: A pulse timed late makes one interval too long and the next too short.
6. 6 Cleaning and gap filling Missed and false pulses are dropped, filled or replaced. Accuracy lost here: Every filled or replaced beat is a guess, and the editing method changes the results.
7. 7 The window and the number Firmware picks a stretch of data, such as short windows at night or at rest, and applies a formula. Accuracy lost here: Short windows hold fewer beats, so a few bad ones weigh more.

Most affected: RMSSD, pNN50, SD1, HF · Less affected: mean HR, LF, SD2

A chain of seven steps from top to bottom: light in and raw signal out, sampling, filtering, finding the same point on every pulse, the interval series, cleaning and gap filling, and the window and the number. Beside each step is the error that can enter there: a weak ripple from the fit, daylight, cold skin or ink, rounding to the sampling ticks, a blunted pulse, noise moving the chosen point, a late pulse lengthening one interval and shortening the next, guessed beats, and short windows. A strip at the end says RMSSD, pNN50, SD1 and HF are the most affected, and mean heart rate, LF and SD2 less so.

1. Light in, raw signal out

An LED, usually green at the wrist, lights the skin, and a photodiode reads the light that returns many times a second. Most of that light comes back unchanged; the beat-by-beat ripple is a small part of it, so the electronics amplify it.

What can go wrong: a loose strap, daylight leaking in, cold skin or dark tattoo ink can all shrink the ripple until it is barely above the sensor’s own noise. PPG explained covers these in detail.

2. Sampling: the clock

The sensor does not see a continuous wave. It takes a reading at fixed ticks, the sampling rate sampling rate Sampling rate How many times a second a sensor measures the signal, in hertz (Hz). It sets how precisely each beat can be timed: about ±1 ms at 1000 Hz, ±4 ms at 250 Hz and ±20 ms at a 25 Hz camera. More about Sampling rate . At 25 samples a second the ticks are 40 ms apart, at 64 they are about 16 ms apart, and at 500 they are 2 ms apart. A pulse can only be placed on a tick, so each pulse time is off by up to half a tick: ±20 ms at 25 Hz and ±1 ms at 500 Hz.

What it costs: every PP interval carries the rounding of the pulse at each end. On its own that error is small next to a 900 ms interval, but HRV is built from changes of 10 to 50 ms, so the rounding can be a large share of what is being measured. From electrode to number works through this with real numbers.

3. Filtering

Before looking for pulses, firmware smooths the signal (a low-pass filter) to cut fast noise, and removes slow drift (a high-pass filter) caused by breathing, posture and pressure on the sensor.

What it costs: smoothing removes noise but also blunts the shape of each pulse, and a filter that is not symmetric in time shifts every pulse by the same amount. A constant shift does not matter for intervals; a shift that depends on the shape of each pulse does.

4. Finding the same point on every pulse

The detector picks one landmark, the fiducial point, on every pulse and times it. There are three common choices:

• The foot, where the pulse starts to rise. Taking the lowest sample is unreliable because the trough is flat. The usual method draws a tangent along the steep upstroke and finds where it meets the level of the trough before it: the intersecting-tangents method.
• The steepest point of the rise, the maximum of the first derivative.
• The peak, the top of the wave.

The diagram below marks all three on one pulse.

1. The foot: where a tangent along the steep upstroke meets the level of the trough before it, the intersecting-tangents method.
2. The steepest point of the rise: noise moves a steep edge very little in time.
3. The peak: near a broad, flat top, a little noise can shift the highest sample by several milliseconds.

One smooth pulse wave. A dashed line runs along the trough before the pulse, and a second dashed line runs along the steep upstroke; where they cross is marked as the foot. A dot halfway up the upstroke marks the steepest rise, with a very short bar across it. A dot on the rounded top marks the peak, with a much wider bar across it, showing that noise can move the peak's time much further.

The steepest rise is usually the most stable, because noise moves a steep edge very little in time. The peak is the least stable: near a broad, flat top, a little noise can shift the highest sample by several milliseconds, and the peak also moves when the shape of the pulse changes with breathing or blood pressure. The evidence agrees. In a tilt-table study that recorded finger and forehead PPG alongside an ECG, the middle of the upstroke, the maximum of the first derivative and the intersecting tangents gave PP intervals closest to the RR intervals, while the peak and the plain foot point agreed least (Peralta and colleagues, 2019). An earlobe study comparing six methods found intersecting tangents the best, with a root-mean-square difference from the RR interval of 5.69 ms (Hemon and Phillips, 2015). The same group of ideas explains a finding from cuffless blood pressure work: pressing the sensor harder shifts the timing of the peak about twice as much as the timing of the foot (Block and colleagues, 2020, citing their earlier study).

Interpolation helps at this step. Instead of using the tick nearest the landmark, the detector fits a small curve through the samples around it and reads the landmark between ticks. For a broad pulse spanning several samples this wins back most of the lost timing.

5. The interval series

Subtract each pulse time from the next and you have the PP series: one number per beat, in milliseconds. Every error from steps 1 to 4 is now inside these numbers. A pulse timed 5 ms late makes one interval 5 ms too long and the next one 5 ms too short.

6. Cleaning and gap filling

Some pulses are missed, during movement or when a premature beat premature beat Ectopic beat A heartbeat that starts outside the heart's normal pacemaker and usually comes early. Ectopic beats are common, but they distort HRV unless removed or corrected. More about Ectopic beat pushes out too little blood. Some false pulses are found, when a step or a twitch looks like a beat. Firmware has to repair the series before it computes anything. Typical rules drop a pulse that comes impossibly soon after the last one, fill a long gap with evenly spaced beats, and replace an interval that differs too much from its neighbours with a local average. HRV standards require abnormal beats and artifacts artifacts Artifact An error in a heartbeat recording, such as a missed, extra or mistimed beat, often caused by movement or poor sensor contact. Artifacts can badly distort HRV numbers. More about Artifact to be dealt with before analysis (Task Force, 1996), and the editing method itself changes the results (Peltola, 2012). Deleting intervals joins beats that were never neighbours; filling them in smooths the series.

What it costs: every filled or replaced beat is a guess, usually a smooth one, so it lowers beat-to-beat variation where it sits. Gaps hurt too. In 20 healthy volunteers, the Apple Watch’s beat intervals had on average five gaps per recording, each about 6.5 seconds long. The time-domain measures were not significantly affected, but LF LF LF power The variance of the heartbeat intervals that falls in the 0.04 to 0.15 Hz band, in ms²: rhythms repeating every 7 to 25 seconds. Once read as sympathetic activity, which is now disputed. More about LF power and HF power HF power HF power The variance of the heartbeat intervals that falls in the 0.15 to 0.40 Hz band, in ms²: the range of normal breathing. It mainly reflects vagal activity when breathing is 9 to 24 breaths a minute. More about HF power were significantly lower (Hernando and colleagues, 2018).

7. The window and the number

Last, firmware picks which stretch of data to use and applies a formula. Most watches and rings do not compute HRV from every beat of the day. They take short windows at night or at rest, average them, and report one value; each brand chooses differently, as HRV by device and the lesson on why devices disagree describe. Short windows hold fewer beats, so a few bad ones weigh more.

Six sources of timing error

Each sketch below matches one of the six causes in the numbered list that follows it.

1. 1 Sampling

   Each pulse is rounded to the nearest tick, up to half the gap between ticks.

2. 2 The landmark

   A broad, flat peak is easy for noise to move; a steep upstroke is not.

3. 3 Transit time

   Breathing, blood pressure and posture change the delay from heart to sensor beat by beat.

4. 4 Movement

   Arm and wrist movement makes swings much larger than the pulse.

5. 5 A weak pulse

   The pulse shrinks while the sensor's noise stays the same, so every landmark jitters more.

6. 6 Missed and extra beats

   Each one forces a repair, and each repair is a guess.

Six small sketches. 1, Sampling: dots on one pulse a few tens of milliseconds apart; none lands on the peak, so the nearest one is off by part of a tick. 2, The landmark: one pulse with a wide span across its flat top and a narrow span across its steep upstroke, showing how far the same noise moves each point in time. 3, Transit time: ECG spikes above pulse waves; the delay from each spike to its pulse is a little different every beat. 4, Movement: a still stretch of small pulses, then swings much larger than the pulse. 5, A weak pulse: a small, noisy pulse under the dashed outline of a strong one, with a wider spread for where each beat is timed. 6, Missed and extra beats: ticks above detected pulses; one small pulse has no tick, and a twitch between two beats gets an extra one.

Each step above adds its own error. Grouped by cause:

1. Sampling. Rounding each pulse to the nearest tick, up to half the gap between ticks. A pilot study that down-sampled PPG recorded at 10,000 samples a second found no measure that differed significantly from ECG HRV above 25 Hz. Below that, pNN50 pNN50 pNN50 The percentage of neighbouring NN intervals that differ by more than 50 milliseconds. It tracks fast, vagally driven changes, like RMSSD, but as a share rather than a size. More about pNN50 broke first (at 20 Hz and below) and RMSSD RMSSD RMSSD Root mean square of successive differences: an HRV metric built from how much each heartbeat interval differs from the next, in milliseconds. It mainly reflects vagal activity. More about RMSSD next (at 15 Hz and below), while mean interval, VLF VLF VLF power The variance of the heartbeat intervals in the 0.003 to 0.04 Hz band, in ms²: rhythms repeating every 25 seconds to 5 minutes. It needs at least a 5-minute recording and is best read over 24 hours. More about VLF power and LF power held up at every rate tested (Choi and Shin, 2017). A study of 57 five-minute recordings drew a stricter line: RMSSD stayed within 5% of the 1,000 Hz value down to 50 samples a second without interpolation, and down to 20 with it; mean pulse rate stayed within 0.05% at 5 samples a second (Béres and Hejjel, 2021).
2. The landmark. A broad, flat peak is easy for noise to move; a steep upstroke is not (Peralta and colleagues, 2019).
3. Transit time. Breathing, blood pressure and posture change the delay from heart to sensor beat by beat (Drinnan and colleagues, 2001; Mukkamala and colleagues, 2015). This adds variability that comes from the vessels. Some researchers see it as a flaw of PRV; others see it as extra information about the circulation (Mejía-Mejía and colleagues, 2020).
4. Movement. Arm and wrist movement shifts the sensor and pushes blood around, producing swings much larger than the pulse. In a study of consumer and research-grade wearables, heart rate error during activity was on average 30% higher than at rest (Bent and colleagues, 2020), and HRV, which needs every beat, suffers far more than heart rate.
5. A weak pulse. Cold hands, a loose fit or low blood flow to the skin shrink the pulse while the sensor’s noise stays the same. Every landmark then jitters more. In a cold-exposure study, PRV from the finger and toe responded differently from ECG HRV, more so than PRV from the ear (Mejía-Mejía and colleagues, Frontiers in Physiology, 2020).
6. Missed and extra beats. Each one forces a repair, and each repair is a guess (Peltola, 2012; Hernando and colleagues, 2018).

Try it: one heartbeat, two sensors

The explorer below is a teaching model of these steps. Five minutes of synthetic resting beats are timed twice: by an ECG, and by a simulated wrist sensor that sees a pulse wave about 200 ms later, samples it, smooths it, finds one point on each pulse and cleans the series. Every HRV parameter is then computed from both, with the same code as this site’s HRV calculator.

ECG and PPG, beat by beat

Synthetic beats, not a recording

Chest strap (ECG) above, watch (PPG) below R peak Timed point Filled in

ECG PPG

0 s 5 s

Delay, R peak to timed point

327 ms

Delay range over 5 min

28 ms

Timing step

±7.8 ms

Beats corrected

0

At 64 Hz, timing the peak, each pulse interval differs from its ECG interval by 7.8 ms (SD). RMSSD reads +11.4% and HF power +20.7%, while mean heart rate is 0.0%.

Sampling rate 64 Hz, one sample every 15.6 ms 255064100128250500

Point timed on each pulse

Foot Steepest rise Peak

Breathing effect on transit time

Off Typical Strong

Movement

Still Light Walking

Pulse strength

Normal Weak (cold hands)

Interpolate between samples

First five beats

Beat	ECG RR ms	PPG PP ms	Difference ms
1	859	859	0
2	863	875	12
3	840	828	−12
4	799	813	14
5	822	813	−9

Five minutes, PPG against ECG

Parameter	ECG	PPG	Difference
Mean HR bpm	68.1	68.1	0.0%
SDNN ms	35.6	36.9	+3.7%
RMSSD ms	30.0	33.4	+11.4%
pNN50 %	9.4	9.1	−3.1%
SD1 ms	21.2	23.6	+11.4%
SD2 ms	45.6	46.5	+1.9%
LF ms²	422	427	+1.1%
HF ms²	334	403	+20.7%
LF/HF	1.26	1.06	−16.2%
DFA α1	0.91	0.87	−4.6%
SampEn	1.97	1.93	−1.7%

Bars run from the centre and stop at ±50%. Amber means the watch reads higher than the ECG, green lower. Amber rows in the beat table were filled in or re-timed by the cleaner.

The controls work when JavaScript is on.

Two traces over the first 5 seconds: an ECG with one tall, narrow R peak per beat, and below it a PPG with one rounded pulse per beat that starts about 200 milliseconds after each R peak. Dashed lines drop from each R peak and short ticks mark the point timed on each pulse. Readouts give the average delay from R peak to that point, how much the delay swings, the timing step of the sensor and how many beats the cleaner corrected. A table lists the first five beat-to-beat intervals from the ECG and the PPG and their difference in milliseconds. A second table lists eleven HRV parameters from five minutes of both, with the PPG's difference as a percentage and a small bar. With the starting settings (64 Hz, the peak, no interpolation, typical breathing, still) the differences are: Mean HR 68.1 on ECG, 68.1 on PPG, 0.0%; SDNN 35.6 on ECG, 36.9 on PPG, +3.7%; RMSSD 30.0 on ECG, 33.4 on PPG, +11.4%; pNN50 9.4 on ECG, 9.1 on PPG, −3.1%; SD1 21.2 on ECG, 23.6 on PPG, +11.4%; SD2 45.6 on ECG, 46.5 on PPG, +1.9%; LF 422 on ECG, 427 on PPG, +1.1%; HF 334 on ECG, 403 on PPG, +20.7%; LF/HF 1.26 on ECG, 1.06 on PPG, −16.2%; DFA α1 0.91 on ECG, 0.87 on PPG, −4.6%; SampEn 1.97 on ECG, 1.93 on PPG, −1.7%. A slider sets the sampling rate from 25 to 500 Hz. Radio buttons choose the point timed on each pulse, the breathing effect on transit time, movement and pulse strength, and a checkbox turns on interpolation between samples.

Things to try:

• Start at 64 Hz on the peak. RMSSD and HF read about a tenth to a fifth high; mean heart rate mean heart rate Mean RR The average time between heartbeats in a recording, in milliseconds. Heart rate is the same number upside down: 60,000 divided by the mean RR. More about Mean RR does not move.
• Drop to 25 Hz. RMSSD jumps by a third or more, pNN50 and HF follow, and DFA α1 DFA α1 DFA α1 The short-term scaling exponent of detrended fluctuation analysis. Near 1 the beats have memory at every scale; near 0.5 they behave like noise. α2 is the same slope over longer stretches. More about DFA α1 falls. Now turn on interpolation and watch most of it come back.
• Switch from the peak to the steepest rise or the foot at 250 Hz. The interval error shrinks.
• Turn breathing to strong with everything else ideal. The transit-time swing alone inflates RMSSD and HF while LF stays put.
• Choose walking. The firmware misses and invents beats, the cleaner guesses, and every beat-to-beat measure falls apart while mean heart rate stays within a few tenths of a percent.

The model is a simplification. Real sensors filter differently, real transit time varies in more complicated ways, and real cleaning rules are private. Treat its numbers as directions and rough sizes, not as a forecast for your device.

Which HRV parameters suffer, and why

The chart below condenses the table later in this section: the more filled segments, the more PPG timing error moves that measure.

Measure Very low High

Time domain

• Mean heart rate : Very low
• SDNN (5 minutes or longer) : Low to moderate (mechanism)
• RMSSD : High (mechanism)
• pNN50 : High, and erratic (mechanism)

Poincaré plot

• SD1 : High (mechanism)
• SD2 : Low (mechanism)

Frequency domain

• HF power : High
• LF power : Low from jitter, moderate from gaps (mechanism)
• LF/HF : Moderate to high (mechanism)

Nonlinear

• DFA α1 : Moderate (mechanism)
• Sample and approximate entropy : Moderate, direction varies (mechanism)

Sensitivity Upper end of a range Not reached

A chart with one row per HRV measure, grouped into time domain, Poincaré plot, frequency domain and nonlinear measures. Each row has a bar of four segments, from very low to high sensitivity, and the table's words for that row.

Two ideas explain most of the table below.

The first is that timing error is mostly random from one pulse to the next. A pulse timed a little late makes one interval longer and the next one shorter, so the error lands squarely on the beat-to-beat differences. If each pulse time is off by a random amount with a standard deviation of σ, each successive difference is off by an error with a variance of 6σ², and RMSSD grows roughly as:

$$\text{RMSSD}_{\text{PPG}}^2 \approx \text{RMSSD}_{\text{ECG}}^2 + 6\sigma^2$$

Read it as: the squared timing noise adds to the squared RMSSD. With σ = 5 ms, 6σ² is 150 ms². A true RMSSD of 30 ms then reads about 32 ms, 8% high; a true RMSSD of 15 ms reads about 19 ms, almost 30% high. The same timing error matters more for people with low HRV. This formula is our derivation from the definition of RMSSD, assuming independent errors, and is a mechanism rather than a published result.

The second idea is about frequency. Random error in pulse times turns into fast, zig-zag error in the intervals, which in a spectrum sits mostly in the high-frequency band, the band where breathing-linked variation lives. Slow rhythms such as the LF band and slow drifts barely see it.

“Mechanism” marks reasoning from how the measure is built rather than a measured result. In practice, firmware cleaning and the choice of window often matter as much as the sensor, so a device can do worse or better than this table predicts.

What the evidence says

Agreement statistics in plain words

Validation studies compare a device with a reference, usually an ECG, on the same people at the same time. Correlation alone is misleading, because two methods can rise and fall together and still disagree by a lot (Bland and Altman, 1986). The usual summary is a Bland–Altman analysis:

• Mean bias is the average difference between the device and the reference. A bias of +3 ms means the device reads 3 ms high on average.
• Limits of agreement are the range that 95% of single differences fall within, the bias plus or minus about two standard deviations. Wide limits mean that even with no average bias, any one reading can be well off.

Many wearable studies also report the mean absolute percentage error (MAPE) and a concordance coefficient (CCC); HRV by device explains those.

At rest and during sleep

Still bodies give the best results.

• A review of studies comparing PRV with HRV found PRV accurate enough at rest, though many studies found short-term variability somewhat overestimated, which the authors traced to breathing acting on the circulation (Schäfer and Vagedes, 2013).
• In 17 young people on a tilt table, PRV and HRV measures did not differ significantly and correlated above 0.97; the differences that remained were mainly in the HF band (Gil and colleagues, 2010).
• PRV worked well as a stand-in for HRV in healthy people at rest, with correlations above 82%, but less well after exercise and in people with heart disease, especially for frequency measures (Pinheiro and colleagues, 2016).
• Overnight, against a chest-strap ECG, in 13 healthy adults over 536 nights: Oura Ring Gen 4 had a CCC of 0.99 and a MAPE of 5.96% for RMSSD, Oura Gen 3 0.97 and 7.15%, WHOOP 4.0 0.94 and 8.17%, Garmin Fenix 6 0.87 and 10.52%, and Polar Grit X Pro 0.82 and 16.32% (Dial and colleagues, 2025).
• In 35 healthy adults at home, an Oura ring measured nocturnal heart rate and RMSSD accurately both in 5-minute windows and as a night average, by the authors’ judgement, although its RMSSD averaged about 15 ms below the ECG’s. SDNN, pNN50 and HF were acceptable only as night averages, and LF and LF/HF had high error either way (Cao and colleagues, 2022).
• In 53 adults in a sleep laboratory, WHOOP 3.0, whose maker provided the raw beat intervals, agreed almost perfectly with the ECG (intraclass correlation 0.99, average underestimate 4.5 ms). Apple Watch, Polar and Oura, using their own processing, had good agreement (0.63 to 0.67), and a Garmin 3-minute test taken while awake had poor agreement (0.24). Several devices read low HRV too high and high HRV too low (Miller and colleagues, 2022).

Strap, ring, wrist and camera side by side

In one study of five young adults over 148 resting trials against a multi-lead ECG, the same app gave a mean RMSSD error of 4.10% with a Polar H10 chest strap and 9.43% with the phone camera. An Oura ring had 6.84%, and another chest-strap system, using a three-minute test, had 11.27% (Stone and colleagues, 2021). In 29 healthy people at rest, smartphone PPG and a chest strap both agreed acceptably with an ECG for RMSSD (Plews and colleagues, 2017). The sensor is one factor among several; recording length and processing count too. Fingertip HRV covers phone cameras, finger clips and rings in more detail.

With movement

Chest straps hold up. Judged against the raw ECG trace, a Polar H10 strap timed 99.6% of beats correctly over activities from easy to hard, and 99.4% during high-intensity exercise (Gilgen-Ammann and colleagues, 2019). Optical sensors do not: physical activity and some mental stressors made PRV agree with HRV much worse, often unacceptably (Schäfer and Vagedes, 2013). That is the main reason most wearables measure HRV during sleep.

The internet debate, fairly

Two camps argue about wearable HRV online.

“PRV is not HRV.” This side is right about the physics. A PPG device measures the pulse at the skin, and that signal carries transit-time changes, vessel tone and pulse shape on top of the heart’s rhythm. Reviews conclude that PRV should not be treated as a valid stand-in for HRV in every situation (Mejía-Mejía and colleagues, 2020), and the gap widens with movement, stress, cold, exercise and heart disease (Schäfer and Vagedes, 2013; Pinheiro and colleagues, 2016).

“Close enough at rest.” This side is right about practice. For a still, healthy person, especially averaged over a night, the best devices land within a few milliseconds of an ECG on average (Dial and colleagues, 2025), and that is small next to the night-to-night changes most people track.

Both can be true at once. The difference matters when:

• you need beat-to-beat precision, such as single short readings, pNN50, HF power or DFA α1;
• you measure during or right after exercise;
• your rhythm is irregular, with frequent extra beats or atrial fibrillation, because those beats are often missed or smoothed away;
• you are doing research, or comparing numbers across devices or with published norms.

It matters much less when you follow your own overnight RMSSD on the same device, worn the same way, night after night. A device that reads 3 ms high every night still shows you which way you are moving.

What to do

• Use one device and compare it only with itself. Each brand has its own sensor, landmark, cleaning and window, so absolute values from two devices are not interchangeable.
• Read trends rather than single values. Overnight averages are more trustworthy than a single short reading.
• Use a chest strap for short morning readings, anything around exercise and research. It times the electrical R peak and holds up during movement.
• Help the optical sensor. Wear it snug, keep still for spot readings, and warm cold hands first.
• Be careful with fine-grained measures from a wearable. Treat pNN50, HF, LF/HF and DFA α1 from a watch as rough.
• If your heartbeat feels irregular, see a doctor. Neither device is a diagnostic ECG.

For the evidence summary in plain terms, read is wrist HRV as accurate as a chest strap?, and for what each brand reports, HRV by device.