Before any explanation, do this. It takes about fifteen seconds and it decides how much of the rest of this page applies to you.
Step one. Open your tracking app and find last night’s heart rate variability number. Write it down.
Step two. Answer one question about where that number came from: was it measured by electrodes touching your skin, or by a light shining into it?
A chest strap reads the electrical signal of your heart. A watch, a ring and a band read light bouncing back off blood moving through tissue. Those two things have different names in the research literature. The first is heart rate variability. The second is pulse rate variability, and a 2026 review by the cardiologist Eric Topol makes the point that almost nobody tells consumers: every study linking this signal to a health outcome used the first kind, and almost every device sold to the public produces the second.
That is not a technicality. One study of over 900 adults compared the two directly and concluded that pulse rate variability is “an invalid surrogate” for heart rate variability — not merely noisier, but underestimating it unevenly, with the size of the gap differing across chronic diseases. A separate systematic review of 43 studies found only 10 of them permitted a pooled quantitative comparison even under ideal conditions, and concluded that optically derived readings “should not be regarded as universally interchangeable” with electrically derived ones.
So the instrument you just used has a name, and it is probably not the name in the research.
Last reviewed: October 2026
Two signals called heart rate variability, and the thresholds their own authors set
Everything below is reported against the bar the researchers themselves declared — their own stated threshold for “acceptable,” “meaningful,” or “powered.” Reading results against their own papers’ criteria, rather than against an outsider’s opinion, is the only way to tell a modest finding from a failed one.
| What the sensor does | What the output is properly called | The authors’ own bar | Verdict against it |
|---|---|---|---|
| Electrodes read the heart’s electrical signal (chest strap, ECG) | Heart rate variability | The reference standard in every outcome study | This is the signal the research is about |
| Light reads blood-volume change in tissue (watch, ring, band) | Pulse rate variability | Agreement holds only when the delay between electrical beat and arriving pulse is constant — roughly, at rest | Acceptable at rest; the two diverge as soon as you are not |
| Pooled comparison, 43 studies | — | Quantitative synthesis possible | Only 10 of 43 qualified, under ideal conditions |
| Direct comparison, >900 adults | — | Correlation good enough to substitute | Failed: “an invalid surrogate,” with error varying by disease |
This matters most for the part of the market that sells certainty. Your overnight reading is a decent estimate at rest, which is exactly when it is taken — so the wrist is not useless. But the chain from your ring’s number to the mortality studies has a broken link in it, and no readiness score mentions that.
There are normal ranges. The range is the finding
The usual line is that there is no good number to aim for. That is the right advice supported by the wrong claim, because reference values do exist, and what they actually show is far more useful than a flat denial.
Nunan and colleagues pooled 44 studies covering 21,438 healthy adults, most aged 40 or over, for short recordings:
| Measure | Mean (SD) | Reported range | The spread, as a multiple | Whose variation this is |
|---|---|---|---|---|
| RMSSD (ms) | 42 (15) | 19–75 | about 4× | between people |
| SDNN (ms) | 50 (16) | 32–93 | about 3× | between people |
| HF power (ms²) | 657 (777) | 83–3,630 | about 44× | between people |
| LF/HF ratio | 2.8 (2.6) | 1.1–11.6 | about 11× | between people |
| Inter-beat interval (ms) | 926 (90) | 785–1,160 | about 1.5× | between people |
Look at the last column against the fourth. Heart rate itself varies by about half again across healthy adults. The beat-to-beat variation varies by four times, and its frequency-domain cousin by forty-four. A single standard deviation on HF power is larger than its own mean.
That is why your number is uninterpretable on its own — not because norms are unavailable, but because the healthy range is so wide that membership in it tells you nothing. A 24-hour study of 2,079 adults aged 25 to 41 found SDNN averaging 160 ± 40 ms in men and 147 ± 36 ms in women, which bears no resemblance to the five-minute figures above; recording length changes the number more than most people’s health ever will.
Two further details that get flattened in general coverage. First, age: time-domain measures do decline with age, but a study of 1,743 adults aged 40 to 100 found RMSSD and pNN50 following a U-shape — falling from 40 to 60, then rising again after 70. Rising beat-to-beat variation in the very old is not recovered youth. Second, sex: a meta-analysis of 296,247 healthy participants found women had lower SDNN but greater HF power than men, so the direction of the sex difference depends on which metric you read. Umetani’s data put the gap before age 30 and found it gone after 50.
What RMSSD is an index of, and what it is not
One metric dominates consumer devices: RMSSD, the root mean square of successive differences. Apple is the notable exception and uses SDNN, which means two people comparing their phones are often comparing different quantities.
RMSSD is described in the reference literature as the primary time-domain estimate of vagally mediated change. The qualifier in that sentence does a lot of work, and the standard overview of these metrics states plainly that “the influence of respiration rate on this index is uncertain.”
| What it is usually said to be | What the physiology literature says | Minimum clean recording its own conventions require |
|---|---|---|
| A measure of parasympathetic tone | HF power “indexes vagal modulation” of heart rate — it “does not represent vagal tone” | HF: 1 minute |
| A number your breathing doesn’t touch | Changing breathing rate anywhere within 9–24 breaths per minute alters HF power without changing mean heart rate | RMSSD: 5 minutes conventionally |
| A clean split between the two nervous-system branches | LF power is “not a pure index” of sympathetic drive; roughly half its variability comes from the parasympathetic side | LF: at least 2 to 2.5 minutes |
| Stable whatever the window | 24-hour and 5-minute LF/HF values correlate poorly | SDNN: 5 minutes; ultra-short proposals of 10, 30 and 60 seconds exist for RMSSD |
The practical consequence is one our breathing exercises page reaches from the other direction: slowing your breath makes the heart-rate oscillation physically larger whether or not your vagal tone has changed. Grossman’s proposed demonstration is to breathe at 6 breaths per minute instead of 15 — the fluctuation grows while mean heart rate holds steady. So an app that rewards you for slow breathing is measuring something real that does not mean what it is taken to mean. Topol notes the matching clinical result: in sham-controlled trials, heart rate variability biofeedback with paced breathing produced no significant effect on stress or on heart rate variability itself.
How accurately does your device measure heart rate variability?
The most-cited recent validation is frequently described as a 536-night study, which makes it sound large. It is 13 adults — six of them women, averaging 33.2 years — each contributing at least ten nights at home, against a Polar H10 single-lead ECG. The nights were treated as independent even though each person supplied roughly forty of them.
| Device | Resting heart rate error vs ECG | HRV error | The paper’s own stated bar | Passes? |
|---|---|---|---|---|
| Oura Generation 3 | 0.98 bpm | ≈ 4 ms | HRV: MAPE above 10% is unacceptable | Yes |
| Oura Generation 4 | 1.08 bpm | ≈ 4 ms | same | Yes |
| WHOOP 4.0 | 1.78 bpm | ≈ 4 ms | same | Yes |
| Garmin Fenix 6 | excluded from the analysis | MAPE 10.52% | same | No |
| Polar Grit X Pro | 1.72 bpm | MAPE 16.32% | same | No |
Read the resting heart rate column and the HRV column as two different verdicts, because they are. Resting heart rate landed within 1.67% to 3.00% of ECG across every device that reported it, comfortably inside the 5-to-7 bpm shift the authors nominated as meaningful. Heart rate variability ranged from 5.96% to 16.32% off, and two of five devices failed the paper’s own acceptability threshold. The accurate number on your wrist is your pulse. The fragile one is the one your recovery score is built on — which is also why our resting heart rate page is the companion to this one rather than a footnote to it.
Two details explain more than the percentages do. Garmin was dropped from the resting-heart-rate comparison entirely because it reports the lowest 30-minute average in a 24-hour period without disclosing when that window occurred, so it could not be time-matched to the ECG at all. And Polar derives its figures from only the first four hours of sleep, so the researchers had to recompute the reference for Polar alone while judging everyone else against the whole night.
The argument the validation started
What happened next is unusually instructive, and it happened in print. Two WHOOP-affiliated researchers published a response arguing the comparison was unfair, because WHOOP weights its heart rate variability toward slow-wave sleep and was being judged against an all-night ECG average.
The original authors replied that they had already run the sleep-stage-matched analysis in house, and WHOOP performed worse under it — so the published comparison was the more generous of the two. They added the reason they validated app-displayed values rather than raw intervals: the app value is what the user sees, and the processing steps behind it are undocumented.
That exchange is the clearest available statement of the real problem. Nobody outside these companies can audit what happens between the sensor and the number.
The error has a direction, and it is up
Here is the part that reverses most people’s instincts about their own data. Noise does not scatter heart rate variability randomly around the truth. It inflates it.
A single ectopic beat — one early or extra contraction, which healthy hearts produce routinely — does this to a clean recording:
| Recording | True RMSSD | With one uncorrected artifact | Multiple | The correction threshold used |
|---|---|---|---|---|
| Full ECG, one ectopic beat | 79 ms | 201 ms | ≈ 2.5× | drop intervals differing >20% from the preceding one (25% in clinical practice) |
| Chest strap during running | 3 ms | 28 ms | ≈ 9× | same rule; usable-data bar set at ≥85% “very good,” ≥80% “good,” <80% “poor” |
| Camera-based optical reading | 51 ms | 137 ms | ≈ 2.7× | same, plus checks for gaps left by removal |
Correction brought those back to 94 ms, 3 ms and 55 ms respectively. The developer who published these examples also notes that interval correction is not performed by the chest-strap sensor itself — a trusted electrode strap hands you uncorrected data and expects your software to clean it.
Invert the usual anxiety. If a drop of 15 points arrives with no explanation, that is within the ordinary night-to-night range of a signal whose healthy spread is fourfold. A sudden spike of 50 or 100 points with no explanation is the reading that deserves suspicion, because that is the shape artifacts make. This is also why a device flagging an irregular rhythm is worth taking to a clinician rather than to a forum — an arrhythmia inflates beat-to-beat variation by construction, and our page on atrial fibrillation covers what that flag does and does not mean.
Where heart rate variability actually varies
The single most clarifying statistic about this measurement is one the consumer conversation never quotes: the intraclass correlation. It partitions the variation in a repeated measurement into the part that is who you are and the part that is how your day went.
A 2025 analysis pooled five longitudinal studies that tracked resting heart rate variability with consumer devices, and reported that statistic for each. Among 717 US knowledge workers wearing a Garmin and measured between midnight and 5 a.m., it came to 0.68 — roughly two-thirds of the variation sitting between people rather than within them. Among 108 German adults with type 1 diabetes on the same device and window, 0.70. Among 25 Dutch student interns taking two-minute morning recordings on a Polar H7, 0.55. Among 55 US adults with a history of head injury taking five-minute morning recordings on a Polar H10, 0.62. And among 525 first-year college students on an Oura ring, averaged weekly, 0.89.
In the largest and longest of those, roughly nine-tenths of the variation in the reading was between people rather than within them. That is the arithmetic behind “don’t compare your number to anyone else’s,” and it is also the arithmetic behind something the advice usually leaves out: if most of the variance is fixed, there is not much left over for yesterday to move.
Which is what the within-person results showed. Across all five studies and dozens of tested associations, five within-person correlations reached significance — and the authors state they applied no correction for multiple comparisons because the work was exploratory. Their own power calculations: about 100 participants to detect a between-person correlation of 0.30, and about 900 observations to detect a within-person correlation of 0.10.
| What was tested | Result | Direction | The convention they declared |
|---|---|---|---|
| Prior-day stress → next-night reading | Non-significant in all five studies (closest: r = −0.08, p = 0.089) | — | r = 0.10 small, 0.30 medium, 0.50 large |
| Prior-day alcohol (n = 25) | r = −0.32, p = 0.001 | as expected | medium-to-large by their own scale |
| Recovery from work / mental exhaustion | r = 0.10, p = 0.004 / r = −0.09, p = 0.001 | as expected | small |
| Positive affect → next night (n = 717) | r = −0.02, p = 0.034 | backwards | below their own threshold for “small” |
| Average stress across people (n = 717) | higher stress went with higher average readings | backwards | — |
| Between-person fatigue / happiness (n = 25) | r = +0.34 / r = −0.37 | both backwards | medium |
| HbA1c / time in target glucose range (n = 108) | r = −0.21, p = 0.014 / r = 0.24, p = 0.019 | as expected | small-to-medium |
| Retinopathy / neuropathy (n = 108) | r = −0.39 / r = −0.28 | as expected | medium |
| Cholesterol, IL-6, IL-10, TNF (n = 108) | all non-significant | — | — |
A correlation of −0.02 reaching statistical significance is not a discovery; it is what 717 people times 25 nights does to a p-value. And note the pattern in the backwards rows: where this signal did move with how people felt, it sometimes moved the wrong way.
The 2025 authors disclose that one is a senior scientist at Gallup and a consultant, and another co-founded a company in which he holds equity.
Does heart rate variability-guided training beat a fixed plan?
Every readiness score makes one implicit promise: train when I say train, rest when I say rest, and you will end up fitter than if you had followed a plan. That is a testable claim, and it has been tested head to head.
A meta-analysis of randomised trials comparing heart rate variability-guided training against a predefined programme found 8 trials and 190 participants in total, from 849 records screened — mostly runners, cyclists and skiers, most programmes about eight weeks.
| Outcome | HRV-guided vs predefined | 95% CI | p | Significant at their declared p < 0.05? |
|---|---|---|---|---|
| VO2max | MD = 0.96 | −1.11 to 3.03 | 0.36 | No |
| Maximal aerobic power/speed | SMD = 0.06 | −0.26 to 0.38 | 0.72 | No |
| Field-test aerobic performance | SMD = 0.14 | −0.22 to 0.51 | 0.45 | No |
| Power/speed at first threshold | SMD = 0.27 | −0.16 to 0.70 | 0.22 | No |
| Power/speed at second threshold | SMD = 0.18 | −0.20 to 0.57 | 0.35 | No |
Five outcomes, five nulls. The authors’ conclusion: guiding training by heart rate variability “did not provide significant benefit” over a fixed plan.
Now look at how the same data can be made to sell the opposite. Within each group, before versus after, the guided group improved VO2max by 2.84 (p < 0.0001) and the fixed-plan group by 1.80 (p = 0.02). Quote only those two lines and the guided approach looks 58% better. The between-group comparison — the only one that answers the question — is the table above. The within-group improvement is the training; the between-group difference is the readiness score, and it is zero.
The authors rate the whole body of evidence at high risk of bias for a reason that cannot be engineered away: no trial blinded participants or assessors, because you cannot conceal from an athlete that their schedule is being set by their own data. Funnel plots showed no asymmetry (Egger p = 0.452 for VO2max), and the paper never reports its I² values. The same review contradicts itself on programme length, saying eight weeks in its results and under eight weeks in its discussion.
What people actually report, and what was measured
A 2026 study followed 39 adults for roughly 85 days each on WHOOP 4.0, matching every night’s RMSSD to a daily yes/no on ten statements. It is the closest thing available to a test of whether the number knows how you feel.
| Daily statement | Reading when “true” | When “false” | p | Matches the assumption? |
|---|---|---|---|---|
| I felt stressed | 63.5 ms | 63.1 ms | 0.63 | No — a 0.4 ms gap |
| I felt nervous or anxious | 62.9 ms | 63.7 ms | 0.41 | No |
| I felt energised | 64.1 ms | 66.9 ms | < 0.01 | Backwards |
| I felt motivated | 63.1 ms | 65.3 ms | < 0.05 | Backwards |
| I drank alcohol | 54.5 ms | 66.2 ms | < 0.001 | Yes — and by far the largest effect |
| I was sufficiently hydrated | 65.4 ms | 59.6 ms | < 0.001 | Yes |
| I felt recovered | 64.5 ms | 61.7 ms | < 0.01 | Yes |
| I was sick or ill | 59.3 ms | 61.9 ms | < 0.05 | Yes |
| I had caffeine | 64.3 ms | 64.4 ms | not reported | No difference |
| I felt emotionally stable | 64.0 ms | 60.7 ms | < 0.05 | Yes |
Stress — the thing most people believe they are tracking — moved the number by four-tenths of a millisecond. Feeling energised and motivated went with lower readings, the reverse of the authors’ hypothesis. The one input that dominates everything else is alcohol, at nearly 12 ms, and the second is hydration, which almost no coverage mentions.
Worth knowing who paid: the study was funded by PepsiCo’s Gatorade Sports Science Institute and every author is a PepsiCo employee. The paper discloses the funding and then declares no conflicts of interest. Its own limitations are frank — small sample, binary responses where a scale would have been better, each item tested separately in a way that “can obscure associations and produce confounded findings.”
What a readiness score is made of
The number you actually look at is not heart rate variability. A review of 14 composite readiness and recovery scores found heart rate variability fed 86% of them, resting heart rate 79%, and physical activity and sleep duration 71% each. Sleep staging — which is itself inconsistent between devices — is folded in on top.
So a readiness score is a blend in which the most accurate ingredient is your pulse, the headline ingredient is the least accurately measured, and the weights are undisclosed and revised by software update. Apple’s Readiness score, announced in September 2026 on a 0-to-10 scale, has no published data behind it anywhere, including on Apple’s own site.
Three studies are commonly cited as outcome evidence for these scores. A study of 389 professional golfers reported that a 10-point rise in Recovery went with about half a stroke fewer per round — no control group, authors employed by the company, and a golf score is not a health outcome. A study of over 2,700 runners is unpublished with no control group or randomisation. A study of around 30,000 users over 72 weeks reported a 5.8 percentage-point self-reported reduction in drinking, again with no control group.
Against that, the strongest causal test available points the other way. In over 46,000 UK Biobank participants, genetically predicted heart rate variability did not reproduce the expected link with mortality — which suggests the signal reflects your physiological state rather than causing your outcomes. Low readings travel with worse health. Nothing yet shows that raising yours improves anything.
What has not been tested
A short list, because the gaps are as informative as the findings:
- No randomised trial has shown that acting on a readiness score improves a health outcome.
- A scoping review found the validation literature largely missing older adults, people of colour (skin pigmentation affects optical sensing), and people who are underweight or obese.
- No standardisation exists across devices, so a 90 on one is not a 90 on another.
- Whether raising your own number changes anything remains untested in humans.
- A hypothesis that this signal might track overnight brain clearance is interesting and unproven.
A note on the numbers in this article
| Claim | What was actually read | Status |
|---|---|---|
| 13 adults, 536 nights, per-device errors, the ≥10% MAPE bar, the Garmin exclusion, the Polar 4-hour window, the published exchange and the reply | A detailed secondary report of the 2025 Physiological Reports validation; the journal page itself would not open | Firm as reported; verify per-device bias and limits of agreement in Tables 1 and 3 of the original |
| Norms, respiration effects, the “does not represent vagal tone” line, recording minimums, the U-shape, the sex meta-analysis | The 2017 Frontiers in Public Health overview, full PDF read | Firm |
| 8 RCTs, 190 participants, all five null between-group results, the within-group figures, risk of bias, the internal inconsistency | The 2020 Applied Sciences meta-analysis, full page read | Firm |
| The five studies, all ICCs, every correlation, the power calculations, the conflicts | The 2025 Sensors analysis, full page read | Firm |
| The ten daily statements and their millisecond values | The 2026 Sensors WHOOP study, full page read | Firm; note its own text/table p-value mismatch on “energised” |
| The artifact inflation figures (79→201, 3→28, 51→137) | The developer’s own published worked examples, read directly | Firm, but they are one person’s recordings, not a sample |
| Pulse rate variability vs heart rate variability, the >900-adult study, the 43-study review, UK Biobank, the 14-score review, the company studies, Apple | A 2026 review article by a cardiologist, read directly | Secondary throughout — each underlying paper should be opened before publishing |
| “A narrative review found HRV useful for training status” | The review exists but was credited to the wrong source in the original, and would not open for this rebuild | Replaced with the randomised-trial evidence above, which addresses the same claim more directly |
Three figures from the original could not be located and were removed rather than re-sourced: the claim of “small-to-moderate” associations with health measures (the within-person associations are mostly null and several significant ones run backwards), and any specific percentage attached to what raises or lowers the number.
When should you see a doctor?
Take a sustained, unexplained decline over weeks to a clinician, particularly alongside fatigue, reduced exercise tolerance or feeling unwell. Take an irregular-rhythm flag, palpitations, dizziness or chest discomfort to one regardless of what any score says — and note that an irregular rhythm inflates this measurement, so a high reading in that context is a reason for concern, not reassurance.
Also worth raising it if checking the number has become compulsive. Sleep-tracking anxiety is a documented phenomenon, and it has a nasty circularity: worrying about the reading is itself an input to the reading.
Frequently asked questions
Is my watch measuring heart rate variability? Strictly, no. Optical sensors measure pulse rate variability. At rest the two track reasonably well, which is why overnight readings are the usable ones — but one direct comparison in over 900 adults called the substitution invalid, and all the outcome research used electrodes.
What is a normal number? Across 21,438 healthy adults, RMSSD averaged 42 ms with a reported range of 19 to 75. That fourfold spread, not the average, is the point.
Does it tell me how stressed I was yesterday? The evidence says no. Prior-day stress was non-significant in all five studies in the 2025 pooled analysis, and a 2026 study found a 0.4 ms difference between nights people called stressful and nights they did not.
Will training by my recovery score make me fitter? Eight randomised trials say it does not beat a fixed plan on any of five performance outcomes.
What actually moves the number? Alcohol, by a wide margin — nearly 12 ms in the 2026 data, larger than every other self-reported input. Hydration and illness follow. Beyond that, most of the variation is between people, not between your days.
Why did it spike for no reason? Probably an artifact. One ectopic beat took a clean 79 ms recording to 201 ms. Unexplained highs are more suspicious than unexplained lows.
The bottom line
Notice who is making the decisions here.
You do not choose which metric your device reports, or whether that metric is RMSSD or SDNN. You do not choose the recording window — Polar takes the first four hours of your sleep, Garmin takes an unspecified half-hour somewhere in the day. You do not choose how much of your score is heart rate variability and how much is step count. You do not get told when the algorithm changes, and it does change. You cannot see whether your readings were artifact-corrected. And the clearest causal evidence available suggests that even if you could move the number, moving it may not move anything else.
What is left for you to decide is small, and it is worth doing well: pick one device and stay on it, watch multi-week trends rather than mornings, log alcohol and illness next to the readings, and treat a surprising high with more suspicion than a surprising low. Then, if the number has started running your mornings, make the one decision that is unambiguously yours — look at it less. On the evidence, you are giving a proprietary composite more authority over your day than its own makers have ever demonstrated it has earned.
This article is for general informational purposes and is not a substitute for personalised medical advice. Tracking of this kind is not a diagnostic tool — consult a healthcare provider about persistent symptoms or irregular heart rhythms. For related context, see our pages on sleep and stress management.
References
- Shaffer F, Ginsberg JP. An overview of heart rate variability metrics and norms. Frontiers in Public Health, 2017;5:258 — the Nunan pooled norms from 44 studies and 21,438 adults; the Aeschbacher 24-hour values; the statement that respiration’s influence on RMSSD is uncertain; that HF power indexes vagal modulation but does not represent vagal tone; Billman’s objection to LF as a sympathetic index; minimum recording lengths; the U-shaped age pattern in RMSSD; and the 296,247-participant sex meta-analysis
- Dial MB, et al. Validation of nocturnal resting heart rate and heart rate variability in consumer wearables. Physiological Reports, 2025;13(16):e70527 — 13 adults, 536 nights, Polar H10 reference, the per-device errors, the >10% MAPE threshold, the Garmin exclusion and the Polar four-hour window
- The Science Behind Wearables — a detailed report of the above validation including the per-device figures, plus the published WHOOP response (Grosicki & Presby, e70710) and the authors’ reply (e70706) stating the sleep-stage-matched analysis was less favourable to WHOOP
- Medellín Ruiz JP, et al. Effectiveness of training prescription guided by heart rate variability versus predefined training. Applied Sciences, 2020;10(23):8532 — 8 RCTs, 190 participants, all five between-group outcomes null, the within-group figures, the high risk of bias from impossible blinding, and the Egger tests
- Hernandez R, et al. Resting heart rate variability measured by consumer wearables and its associations with diverse health domains in five longitudinal studies. Sensors, 2025;25(23):7147 — the five cohorts, all intraclass correlations, every within- and between-person correlation, the power calculations, the absence of multiple-comparison correction, and the authors’ disclosures
- Ungaro CT, et al. Disconnection between self-reported wellbeing and heart rate variability from wearables. Sensors, 2026;26(4):1325 — 39 adults, roughly 85 days each on WHOOP 4.0, the ten daily statements with their millisecond values, the backwards results for “energised” and “motivated,” the alcohol and hydration effects, and the PepsiCo funding
- Topol E. The big holes in wearable heart rate variability and readiness scores. Ground Truths, September 2026 — the pulse rate variability distinction, the >900-adult study calling it an invalid surrogate, the 43-study systematic review, the UK Biobank Mendelian randomisation in over 46,000 participants, the review of 14 composite scores, the sham-controlled biofeedback nulls, the company-funded golfer, runner and alcohol studies, and Apple’s unpublished Readiness score
- Altini M. Issues in HRV analysis: motion artifacts and ectopic beats — the worked examples in which one ectopic beat moved RMSSD from 79 to 201 ms, a running artifact from 3 to 28 ms, and an optical artifact from 51 to 137 ms; the 20% correction rule; the usable-data thresholds; and the note that chest-strap sensors do not correct intervals themselves
