How to read an HRV study or a device validation

An HRV paper is only as good as what it tells you about the chain from lesson 16 and the settings from lessons 17 to 19. Two published checklists say what a paper must state. A device validation adds four statistics, and each one answers a different question. This lesson shows how to read both, with a 2025 wearable study as the worked example, and ends with what the word “validated” can and cannot mean on a box.

Part 1: can the study be reproduced?

Start with the methods section, not the abstract. If you cannot tell how the beats were recorded, cleaned and analysed, the result cannot be compared with any other, whatever it claims.

The GRAPH checklist (Quintana, Alvares and Heathers, 2016) was written for psychiatry but applies to any HRV study. It has 13 items in four domains. Participant selection: how each group was recruited and assessed, the inclusion criteria, disease characteristics and medication, and demographics including physical activity, alcohol and nicotine. Interbeat interval Interbeat 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 collection: the hardware and electrode configuration, the raw 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 , recording length, time of day, filtering, posture and instructions. Data preparation: how intervals were calculated and resampled, how 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 were identified, why data were lost, and how they were cleaned, with the percentage of beats corrected. HRV calculation: the metrics, software and any log transform, and the frequency bands with their interpretation. The paper notes that a 5% data-loss cutoff is common and has no empirical basis, and that FFT and AR estimates of HF HF 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 correlate at r = 0.96, so the method matters less than the reporting of it.

The Catai checklist (Catai and colleagues, 2020) has 30 items aimed at cardiology and cardiac rehabilitation and goes further into conditions: the room (20 to 24 °C, 40 to 60% humidity, noise, people passing), at least 15 minutes of rest in the recording posture, substances avoided the day before and on the day (alcohol, caffeine, nicotine, chocolate, soft drinks, energy drinks), no strenuous exercise the day before, the time of day, breathing rate breathing rate Breathing rate How many breaths you take a minute. It moves HRV on its own: slower, deeper breathing makes the beat-to-beat swing larger, and below about 9 breaths a minute it shifts from the HF band into LF. More about Breathing rate checked (below 8 breaths a minute, the 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 band cannot be interpreted), a stated signal-quality criterion, the number of beats omitted and interpolated, and whether windows were set in time or in beats.

The Psychophysiology guidelines (Quigley and colleagues, 2024) add the hardware floor: 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 at 1000 Hz as the default, 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 and rates down to 100 Hz only where ECG is not feasible, the number of beats edited, any gap over 5 seconds, a test of the editing’s effect whenever it exceeds 1%, and for consumer devices, a validation against a reference.

Read a paper with these three lists open. Heathers (2014) surveyed 97 short-term HRV studies published in 2012 and found that only 45% controlled for food, 23% for water and 4% for a full bladder before recording. A study that reports a group difference in HF without the breathing rate has reported a difference it cannot interpret (lesson 18).

Go deeper Design questions the checklists do not ask

Laborde and colleagues (2017) frame a study around three Rs: resting (the level at baseline), reactivity (the change from baseline to a task) and recovery (the change after it). A paper that measures only one of the three has answered only a third of the question, and the three are not interchangeable: higher resting vagal tone is usually good, while a vagal withdrawal under load can be adaptive or not depending on the task. Ask which R the study measured, whether the baseline posture matched the task posture, and whether the person cleaning the beats knew which condition they were cleaning. Laborde recommends they should not.

Part 2: the four statistics of agreement

A device validation compares a new instrument with a reference on the same beats. Four numbers appear, and they answer different questions.

Correlation does not measure agreement. Pearson’s r says whether two measures rise and fall together. A watch that reads exactly double the ECG scores r = 1. Bland and Altman (1986) opened with that point, “the use of correlation is misleading”, and gave the method that replaced it.

Bland-Altman: bias and limits of agreement. For each pair of measurements take the difference $d$ and plot it against the pair’s mean. The average difference is the bias, and the limits of agreement are

$$\bar{d} \pm 1.96\,s_d$$

Read it as: the mean difference, plus and minus about two standard deviations of the differences. The 1986 paper rounds 1.96 to 2. If the differences are roughly normal, 95% of single comparisons fall inside those limits. The bias tells you whether the device reads high or low on average; the limits tell you how far one reading can be off. A validation that reports the bias without the limits has reported the average and hidden the scatter.

ICC: the intraclass correlation coefficient. This treats the two instruments as raters and asks what share of the total variance comes from real differences between people rather than from disagreement between raters:

$$\mathrm{ICC} = \frac{\sigma^2_{\text{between people}}}{\sigma^2_{\text{between people}} + \sigma^2_{\text{error}}}$$

Read it as: the variance that is signal, divided by signal plus noise. Koo and Li (2016) describe ten forms of ICC, each with different assumptions, and ask authors to state the model, the type and the definition they used, and to report the 95% confidence interval. Their bands: below 0.5 poor, 0.5 to 0.75 moderate, 0.75 to 0.9 good, above 0.9 excellent. An ICC depends on how different the people in the sample are: a spread-out sample makes a device look more reliable than the same device on similar people.

CCC: the concordance correlation coefficient. Lin (1989) defined it as the correlation between the two readings along the 45-degree line through the origin. It multiplies Pearson’s r (precision, how tight the cloud is) by a bias correction factor (accuracy, how far the cloud sits from the line of perfect agreement):

$$\rho_c = \frac{2\rho\,\sigma_x\sigma_y}{\sigma_x^2 + \sigma_y^2 + (\mu_x - \mu_y)^2}$$

Read it as: twice the covariance, divided by the two variances plus the squared difference between the two means. A constant offset or a scale error lowers it; a tight cloud on the line gives 1. It is the coefficient the wearable study below uses.

MAPE: mean absolute percentage error. The average of the absolute differences, each as a percentage of the reference:

$$\mathrm{MAPE} = \frac{100}{n}\sum_{i=1}^{n}\left|\frac{x_i - r_i}{r_i}\right|$$

Read it as: on average, how many percent off. It is easy to read and it has a trap: for a measure like 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 , which can be 15 ms in one person and 90 ms in another, the same 5 ms error is 33% for one and 6% for the other.

Part 3: a worked example

Dial and colleagues (2025) had 13 healthy adults (6 women, mean age 33) wear a Polar H10 chest strap, a single-lead ECG sampled at 1000 Hz, alongside five wearables for 536 nights in all, and compared nightly RMSSD. The windows differed: the rings and the Garmin averaged 5-minute segments across the night, WHOOP used an average weighted towards the last stretch of slow-wave sleep, and the Polar watch used the four hours after sleep onset. The reference was the strap’s whole-night value from Kubios, kept only when under 5% of beats needed correction.

The paper gives the limits to two decimals; they are rounded to one here. Nights per device ran from 139 (Oura Gen 4) to 470 (Oura Gen 3), and every bias in the paper’s table is negative: −1.0 ms for the Gen 4 ring, −2.5 for the Gen 3, −0.8 for WHOOP, −1.8 for the Garmin and −4.7 for the Polar watch. For four devices the limits are the bias plus and minus 1.96 standard deviations, as they should be. The Polar row is not: its bias is −4.65 ms with a standard deviation of 9.67 ms, which gives limits of −23.6 to +14.3 ms, the printed limits with their signs swapped. One of the two signs in that row is a misprint, and the paper does not say which. Either way, the Polar watch’s band is about 38 ms wide, against about 18 to 23 ms for the rings and WHOOP.

The authors graded CCC as under 0.80 unacceptable, 0.80 to 0.89 poor, 0.90 to 0.95 moderate, 0.95 to 0.99 substantial and above 0.99 nearly perfect, and treated a MAPE above 10% as unacceptable. Read the table through the four statistics. Every device has a small bias: on average they all read within a few milliseconds of the strap. The limits are what separate them: the rings and WHOOP keep 95% of nights within about 12 ms either way, while the Polar watch’s nights spread across a band nearly twice as wide. The CCC ranks them in that order because it punishes the spread, and the MAPE puts two devices past the authors’ line. The study’s own caveats: 13 people, all apparently healthy, and proprietary algorithms that may be updated, which is why the authors ask for continued revalidation.

Part 4: group result, individual number

The table describes devices across people. Your question is different: is this device precise enough to show the changes you care about in your own data? That depends on a number no validation can give you, the size of your own day-to-day scatter.

Show 7-day averageShow my usual range

RMSSD, ms

30405060

29 days agoToday

• One morning's reading
• 7-day average
• Usual range

With nothing going on, single readings still swing between 37 and 57 ms. The 7-day average stays between 45 and 49 ms.

Add something that happens

A bad nightTwo drinksA hard training weekGetting illOne reading taken standingReset

Simulated morning RMSSD for the last 30 days. Usual range 43 to 51 ms.

Day	Reading (ms)	7-day average (ms)	Event
29 days ago	41	48	None
28 days ago	43	47	None
27 days ago	47	46	None
26 days ago	45	45	None
25 days ago	46	45	None
24 days ago	52	46	None
23 days ago	51	46	None
22 days ago	45	47	None
21 days ago	52	48	None
20 days ago	37	47	None
19 days ago	42	46	None
18 days ago	40	45	None
17 days ago	52	45	None
16 days ago	57	46	None
15 days ago	53	47	None
14 days ago	42	46	None
13 days ago	41	46	None
12 days ago	46	47	None
11 days ago	53	49	None
10 days ago	46	48	None
9 days ago	43	46	None
8 days ago	45	45	None
7 days ago	51	46	None
6 days ago	46	47	None
5 days ago	47	47	None
4 days ago	52	47	None
3 days ago	41	46	None
2 days ago	40	46	None
Yesterday	45	46	None
Today	52	46	None

A chart of 30 simulated morning RMSSD readings that bounce up and down from day to day, with a steadier line for the 7-day average and a shaded band for the usual range. A table with every reading follows the chart.

Suppose your mornings spread ±20 ms around a 7-day average of 60 ms. A device whose single nights sit within about ±12 ms of the truth adds less noise than you already have, and its trend is usable. Suppose your scatter is ±5 ms around 25 ms. The same device’s error is now larger than your own variation, and a one-night change of 8 ms is as likely to be the sensor as you. The remedy is the one from lesson 10: read the 7-day average, not the night. A bias matters less than it looks, because a device that always reads 2.5 ms low still shows your changes faithfully; it matters only when you compare your number with someone else’s, or with a norm table built on another device (lesson 7).

The same logic applies to every group finding in this course. A meta-analysis that links low HRV with cardiovascular events (lesson 14) describes a population. An ICC of 0.9 describes how well a device separates people from one another. Neither is a statement about one reading on one morning.

Part 5: reading a risk headline

Studies of HRV and health end in a different kind of number: a relative risk. “People in the lowest HRV group had a 35% higher risk” means the event was 1.35 times as common in that group as in the comparison group. It does not say how common the event was to begin with, and that, the absolute risk, is what decides how many people the headline is about.

The widget below turns a relative risk into people. Set a baseline baseline Baseline Your own usual HRV level and range, built from weeks of readings taken the same way. Comparing a reading with your baseline says more than comparing it with other people. More about Baseline , how many people in 100 would have the event anyway; the baseline here is a made-up example, not anyone’s real risk. Then load a relative risk from one of three studies cited on the HRV and longevity page: a meta-analysis of first cardiovascular events in people without known heart disease (Hillebrand 2013), a meta-analysis of death from any cause (Jarczok 2022), and a 1987 study of survivors of a heart attack (Kleiger). Count the extra people each one means.

Relative risk, in people

A made-up baseline, real relative risks

Baseline risk 4 in 100 Relative risk 1.35 times the risk

Hillebrand 2013: 1.35 Jarczok 2022: 1.56 Kleiger 1987: 5.3

At baseline

4 in 100

Extra people

1 in 100

• Event at baseline
• Extra at the higher risk
• No event either way

With 4 in 100 at baseline, a relative risk of 1.35 means about 5 in 100: 1 more person, and 95 of the 100 with no event either way.

The sliders work when JavaScript is on.

A ten-by-ten grid of 100 circles. Filled amber circles are the people who would have the event at the baseline risk, amber-outlined circles are the extra people at the higher relative risk, and grey circles are people with no event either way. One slider sets the baseline risk from 1 to 30 in 100; another sets the relative risk from 1 to 6. Three buttons load the relative risks from Hillebrand 2013, Jarczok 2022 and Kleiger 1987. A sentence counts the people in each group.

Read any risk headline with three questions. Relative to whom: the highest HRV group, the rest of the sample, or a fixed cut-off? Out of how many: what is the absolute risk in the comparison group? And in whom: healthy adults, or patients after a heart attack, measured with a medical ECG over 24 hours rather than a wearable overnight? A relative risk of 5 after a heart attack and one of 1.35 in healthy adults are different findings about different people, and neither is a forecast for one person.

Part 6: what “validated” means on a box

“Clinically validated” is not a regulated phrase for consumer HRV. When you meet it, ask six questions, and expect the answer to each to be in a published paper, not on the product page.

1. Against what? A 1000 Hz ECG or chest strap is a reference; another wearable is not.
2. On whom? How many people, what ages, how healthy, and whether anyone with an irregular rhythm was included. The 2025 study is 13 adults with a mean age of 33.
3. Which metric and window? Nightly RMSSD from 5-minute segments is not the same quantity as a 5-minute morning reading or a 24-hour SDNN SDNN SDNN The standard deviation of all NN intervals in a recording, in milliseconds. It measures overall variability, and its value depends strongly on how long the recording lasts. More about SDNN (lesson 17).
4. Which statistic? Correlation alone is not agreement. Look for bias with limits, and an ICC or CCC with its interval.
5. Which firmware? Algorithms change. A validation of last year’s firmware is a historical document.
6. Who paid? Maker-funded studies can be sound, and should say so.

A device that passes all six is validated for that use. It is still not a diagnostic instrument, and lesson 14’s red flags still apply whatever its CCC.

You have reached the end of Level 4 and of the course. The certificate page checks every quiz, all twenty, and prints your certificate once they are passed.