How to study HRV: a guide for researchers

This guide is for anyone planning a study or trial with heart rate variability (HRV) as a measure: a lab stress task, an exam, or days of continuous recording in daily life. It ends with a checklist you can work through before, during and after data collection. Reading HRV papers is covered in the lesson how to read an HRV study; this page is about running one.

What HRV can and cannot index

HRV is the variation in the time between heartbeats. The measures most stress studies use come from short recordings and mainly reflect the vagus nerve vagus nerve Vagus nerve The main nerve of the parasympathetic nervous system, running from the brainstem to the heart, lungs and gut. Its fast signals to the heart drive most beat-to-beat HRV. More about Vagus nerve , the parasympathetic parasympathetic Parasympathetic nervous system The "rest and digest" branch of the autonomic nervous system. It slows the heart, mainly through the vagus nerve, and drives most of the beat-to-beat variation in HRV. More about Parasympathetic nervous system “brake” that slows the heart:

• 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 , the root mean square of successive differences: how much each gap between beats differs from the next, in milliseconds.
• lnRMSSD lnRMSSD lnRMSSD The natural logarithm of RMSSD, with no unit. Taking the log evens out the spread of RMSSD values, which makes day-to-day changes easier to compare and analyse. More about lnRMSSD , the natural logarithm of RMSSD, which makes its skewed distribution closer to normal.
• 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 , the variability at breathing frequencies, 0.15 to 0.4 Hz in the Task Force definition (Task Force, 1996).
• Respiratory sinus arrhythmia Respiratory sinus arrhythmia Respiratory sinus arrhythmia The normal rise in heart rate as you breathe in and fall as you breathe out. It is driven by the vagus nerve and makes up much of resting HRV. More about Respiratory sinus arrhythmia (RSA), the speed-up of the heart on each breath in and the slow-down on each breath out, which HF power and RMSSD largely capture.

Laborde and colleagues (2017) treat these as measures of cardiac vagal tone cardiac vagal tone Vagal tone The ongoing influence of the vagus nerve on the heart, which slows it down. It is usually estimated from HRV measures such as RMSSD and HF power. More about Vagal tone and recommend analysing one main one, RMSSD, a peak-to-valley RSA measure or HF power. Shaffer and Ginsberg (2017) give the background for each.

What HRV does not index is stress itself. A 2018 review of 37 studies found that HRV changed in response to stress in most of them, and the most common finding was lower parasympathetic activity (Kim and colleagues, 2018). The same drop follows other loads. In about 9 million morning readings from a phone app, HRV fell after alcohol, hard training and sickness, which led the authors to call it a sensitive but not specific marker of stress (Altini and Plews, 2021).

Four things move HRV regardless of stress, and a design has to deal with each:

• Breathing. The rate and depth of breathing change RSA and HF power, which can confound them as measures of vagal tone (Grossman and Taylor, 2007).
• Posture and movement. Standing up, talking and walking change HRV. Grossman and Taylor describe RSA as “profoundly influenced” by momentary physical activity, which can bias estimates of vagal tone.
• Heart rate. Every common HRV measure rises as the gap between beats gets longer, so a faster heart tends to show less variability. De Geus and colleagues (2019) argue that HRV is neither simply nor uniformly a stand-in for heart rate, and close with guidance on reporting the two together.
• The LF/HF ratio LF/HF ratio LF/HF ratio LF power divided by HF power, with no unit. It was proposed as a gauge of balance between the sympathetic and parasympathetic systems, but research shows it cannot measure that reliably. More about LF/HF ratio . It does not measure a “balance” between the sympathetic sympathetic Sympathetic nervous system The "fight or flight" branch of the autonomic nervous system. It speeds up the heart and makes it pump harder, but it acts too slowly to drive beat-to-beat HRV. More about Sympathetic nervous system and parasympathetic branches (Billman, 2013). Avoid it as a stress index stress index Baevsky stress index A number built from the histogram of heartbeat intervals, with no unit. It rises as the rhythm gets steadier and less variable, which Baevsky read as strain on the body's regulation. More about Baevsky stress index .

Three study designs

All three designs share what Laborde and colleagues (2017) call the three Rs of HRV: resting (the 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 ), reactivity (the change from baseline to the event) and recovery (the change from the event to afterwards).

A lab stressor: baseline, task, recovery

The most controlled design brings people into the lab, records a resting baseline, applies a stressor and records recovery. The best known stressor is the Trier Social Stress Test (TSST): about 10 minutes of preparation, then 10 minutes of a free speech and mental arithmetic in front of an audience. Across six studies it raised heart rate and lifted salivary cortisol two- to fourfold above baseline (Kirschbaum and colleagues, 1993). Simpler tasks are also used, such as timed arithmetic alone or computer work with added mental stressors. In one such study, 12 women did the same computer work twice, once with stressors added. HF power was lower in the stressful session than in the control session and partly recovered during 8 minutes of rest, while blood pressure did not get back to its resting level in that time (Hjortskov and colleagues, 2004).

1. Settle at least 5 min
2. Baseline 5 min
3. Prepare 10 min
4. Speech and arithmetic 10 min
5. Recovery at least 10 min
6. Resting
7. Reactivity
8. Recovery

A timeline of five phases sized by their length: settle for at least 5 minutes, baseline for 5 minutes, prepare for 10 minutes, speech and arithmetic for 10 minutes, and recovery for at least 10 minutes. A second row groups them: resting covers settle and baseline, reactivity covers prepare and the speech and arithmetic, and recovery covers the last phase.

How long each phase should be:

• Settle. Laborde and colleagues (2017) ask that people hold the test posture for at least 5 minutes before the baseline starts. Seated with the knees at a right angle is the usual default.
• Baseline. Five minutes is the short-term standard (Task Force, 1996). Jennings and colleagues (1992), who measured heart rate and blood pressure in 48 young men, note that even 20 minutes of rest does not guarantee a stable baseline, and judged 10-minute resting baselines acceptably stable. Their “vanilla” baseline, a simple colour-detection task with little demand, did as well as plain rest or better.
• Task windows. A 1-minute lnRMSSD has good reliability (Laborde and colleagues, 2017). HF power needs at least about 1 minute and LF power LF power 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 about 2 (Task Force, 1996). Compare windows of the same length: values from recordings of different lengths are not interchangeable (Shaffer and Ginsberg, 2017).
• Recovery. No guideline fixes its length, so make it long enough to see the return and state the length in advance. Blood pressure can take longer than HRV: in Hjortskov’s study, HRV had partly recovered after 8 minutes of rest, but blood pressure had not returned to resting values.

Talking changes breathing, so HRV during a speech partly reflects the speech; record breathing and read the speech window with care. A control session with the same timing and posture but no stressor, in counterbalanced order, separates the stressor’s effect from the effect of sitting in a lab for 40 minutes.

A real-life stressor: exams

Exams are a stressor students face anyway, with no manipulation by the researcher. You cannot randomise who sits an exam or when, so each student serves as their own control: measure the same person at the same time of day, in the same posture, on days with and without the exam.

1. Weeks before Resting baseline at a set time and posture
2. Day before Evening reading, then the night’s sleep
3. Exam day Before, during if possible, and after
4. Days after Recovery readings
5. Weeks after A second baseline, away from exams

A row of five points on a line: weeks before, a resting baseline at a set time and posture; day before, an evening reading then the night's sleep; exam day, highlighted, with readings before, during if possible, and after; days after, recovery readings; weeks after, a second baseline away from exams.

Each point answers a different question. The baseline weeks before is the reference. The evening and night before show anticipation and how the person slept, and sleep moves HRV on its own. Readings just before, during and after the exam give reactivity and recovery. A second baseline weeks later, away from exams, separates the exam from slow changes across a term or a season, since a single baseline only allows a before-and-after comparison.

Three studies show the range of designs. Lucini and colleagues (2002) examined university students shortly before an exam and again 3 months later, during the holidays. On the exam day the gap between beats was shorter, blood pressure higher, a normalised low-frequency measure higher and baroreflex gain baroreflex gain Baroreflex The reflex that steadies blood pressure from beat to beat by changing heart rate. Its built-in delay makes heart rate rise and fall about every 10 seconds, the 0.1 Hz rhythm in LF power. More about Baroreflex lower, and salivary cortisol was 5.6 against 2.4 ng/mL. Dimitriev and colleagues (2016) recorded 96 students lying down for 5 minutes on two different days, once at rest and once just before an oral exam. Before the exam, heart rate averaged 83 against 72 beats a minute, and SD1 SD1 SD1 and SD2 Two measures of the spread of dots in a Poincaré plot, in milliseconds. SD1 is the width and tracks short-term variability, like RMSSD; SD2 is the length and tracks overall variability. More about SD1 and SD2 , the Poincaré counterpart of RMSSD, 26 against 36 ms. The drop in SD1 held after adjusting for heart rate, while the drops in 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 and HF power did not. Hammoud and colleagues (2018) monitored 90 students for 4 hours, 1 before, 2 during and 1 after a final exam, with no recording on a calmer day. SDNN was highest in the hour after the exam, and RMSSD was higher after the exam than during it. Heart rate was highest in the hour before the exam, spent on routine pre-exam activities, and lowest during it.

Days of continuous recording with event markers and diaries

The most realistic design records every beat for one or more days while people go about their lives. Events are logged in two ways: event markers, a button press or app entry at the moment something happens, and ecological momentary assessment (EMA), short questionnaires on a phone several times a day about mood, worry, activity or caffeine.

Beats from an ECG strap

Movement

Event markers

Diary prompts

Sleep

Day 1Day 2Day 3

Five lanes across three days with the nights shaded. The beats lane is a continuous line, higher during the nights. The movement lane has bars of varying height during the day and none at night. The event markers lane has a few vertical marks each day. The diary lane has five dots each day, about every three hours from morning to evening. The sleep lane has a block each night.

Two studies show what this design can find. Brosschot and colleagues (2007) gave 52 healthy people an hourly diary for one day while an ambulatory 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 recorded their heart. Stressors and worry went with higher heart rate and lower HRV while awake, the effects carried into the following night’s sleep, and worry duration mediated the effect of stressors. Verkuil and colleagues (2016) went a step further with 32 people: a short calibration learned how each person’s RMSSD fell with movement, and the next 24 hours were searched for periods of at least 7.5 minutes in which RMSSD fell more than movement explained. Those periods went with worry, less positive mood and more tension, but not with the number of stressful events reported.

Movement is the main problem with daily-life data. Physical activity changes RSA a great deal (Grossman and Taylor, 2007), so a design that records movement can tell a stress response from a walk, and one that does not may end up measuring how active people were.

The measurement chain

Every HRV number in a study passes through the same chain. An error at any step carries into the statistics.

1. 1 Sensor ECG chest strap or Holter; PPG watch or ring
2. 2 RR intervals The time between beats, in ms
3. 3 Artifact correction Find, fix and count the bad beats
4. 4 Windows Under 5 min, 5 min, or 24 h
5. 5 Metrics RMSSD, lnRMSSD, HF power
6. 6 Statistics Within-person change, mixed models

Hold steady or record breathing, posture, time of day, caffeine, alcohol, recent exercise, menstrual cycle phase, medication

Six numbered steps in a row: sensor, an ECG chest strap or Holter, or a PPG watch or ring; RR intervals, the time between beats in milliseconds; artifact correction, finding, fixing and counting the bad beats; windows, under 5 minutes, 5 minutes or 24 hours; metrics, RMSSD, lnRMSSD and HF power; statistics, within-person change and mixed models. Under them, a band lists what to hold steady or record: breathing, posture, time of day, caffeine, alcohol, recent exercise, menstrual cycle phase and medication.

Sensor

An ECG times each beat from the sharp R peak of the heart’s electrical signal and is the reference for HRV; the Task Force (1996) gives 250 to 500 samples a second as the optimal range. A chest strap is a single-lead ECG and works well for lab sessions and days of recording; a clinical Holter adds leads and beat-by-beat review. Watches and rings use photoplethysmography photoplethysmography 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 (PPG), which times the pulse wave in the skin instead. A review of studies comparing the two found the pulse accurate enough at rest, but physical activity and some mental stressors made the agreement worse, often to an unacceptable degree (Schäfer and Vagedes, 2013). For a stress task that is a poor fit; what PPG does to HRV numbers covers the errors step by step.

RR intervals and artifact correction

The sensor gives a series of RR intervals RR intervals 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 , the times between successive beats. Missed beats, extra beats and ectopic beats ectopic beats 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 (early beats that do not start in the heart’s natural 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 ) distort the measures, and short recordings are more sensitive to them than long ones. Editing methods delete, interpolate or filter the bad beats, and the choice changes the HRV values; how much depends on the setting, the method, the measure, the people and the length of the recording (Peltola, 2012). Choose the method and an exclusion threshold for the share of corrected beats corrected beats Beat correction Finding the beat intervals that are wrong (missed, extra or ectopic beats) and deleting or replacing them before HRV is calculated. The rule of thumb is to correct under 5% of beats. More about Beat correction before you see the outcomes, and report the share corrected in every recording. Artifacts and extra beats shows what the errors look like.

Windows

• Ultra-short, under 5 minutes. In 3,387 adults at rest, a single 10-second ECG already gave a valid RMSSD, and 2 minutes agreed with 4 to 5 minutes almost perfectly; RMSSD held up better than SDNN at every length (Munoz and colleagues, 2015). That was at rest; a task window still has to fit inside the task.
• Short, the 5-minute standard (Task Force, 1996).
• Long, a nominal 24 hours, which captures the day-night rhythm day-night rhythm Circadian rhythm The body's 24-hour clock. Heart rate is lower and vagal HRV higher at night than by day, so readings are only comparable when they are taken at the same time of day. More about Circadian rhythm and is used for different measures (Task Force, 1996).

Keep one window length per comparison.

Metrics

For short windows, time-domain measures such as RMSSD are the safer choice. RMSSD was valid from 10 seconds at rest (Munoz and colleagues, 2015), while HF power needs about a minute (Task Force, 1996), and RMSSD is less tied to breathing than HF power (Laborde and colleagues, 2017). HF power assumes the 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 stays inside its band: below about 9 breaths a minute (0.15 Hz) the breathing wave moves into the LF band, which is how slow breathing inflates LF power. The lesson frequency analysis without the fog works through it, and RMSSD, SDNN and lnRMSSD explains the time-domain measures.

Controls

Laborde and colleagues (2017) list what to hold steady or record before a session:

• Sleep. A normal night’s sleep before the session, with usual bed and waking times recorded.
• Exercise. No intense training the day before. After exercise the vagal brake returns in stages, over the first 90 minutes and then over up to 48 hours (Stanley and colleagues, 2013).
• Food and drink. No meal, coffee, tea or other caffeinated drinks in the 2 hours before, and no alcohol for 24 hours.
• Time of day. The same time for every session and every participant.
• Posture. The same posture throughout, held for at least 5 minutes before the baseline.
• Breathing. Record it rather than pace it. Laborde and colleagues advise against routinely correcting HRV for breathing when people breathe at their own pace. They note that paced breathing can shift HF power in ways that are hard to predict and can get in the way of a task, and they ask researchers to check that breathing stays between 9 and 24 breaths a minute, the HF band.
• Medication. Record any medication that acts on the heart.
• Menstrual cycle. A meta-analysis of 37 studies and 1,004 people found vagal activity lower in the luteal than in the follicular phase (d = −0.39) and recommends recording or controlling cycle phase (Schmalenberger and colleagues, 2019).

Every later number is compared with the resting baseline these rules produce, so hold the baseline to the same standard as the task.

Continuous HRV with event markers

A day of 5-minute windows is hard to read as it stands, because everything that happened that day is in it. Event-locked averaging lines up the windows around every marker of one kind, from some time before the marker to some time after, and averages them. Whatever the event does to HRV adds up across events; whatever happens at random times averages out. It is the same idea as an averaged evoked response in EEG, applied to heart rate.

One recording, three markers

Lined up on the marker and averaged

Before Marker After

Average

Two charts. The first is a wavy line over six hours with three vertical marker lines; after each marker the line dips and slowly recovers, and a shaded band covers 30 minutes before to 90 minutes after each marker. The second chart has a vertical marker line a quarter of the way across, with a shaded area before it, three thin lines and one thick line. All the lines are level before the marker, drop just after it and climb back over the rest of the chart. Labels read before, marker and after.

Try it on three made-up days of a student in exam week. The whole-day view shows heart rate and RMSSD with every exam, walk and coffee marked. Choose a marker to lock the windows to it.

Lock the windows to the markers

A teaching model, not a recording

Show

Whole days Exam start Exam end Walk Coffee

RMSSD in 5-minute windows (ms) Exam Walk Coffee

204060

Heart rate (bpm)

6080100

Day 1Day 2Day 3

Exam windows

28 ms

All other windows

38 ms

Other windows, walks left out

42 ms

Heart rate and RMSSD, r

−0.89

Over the three days, exam windows average 28 ms against 38 ms for all other windows, 27% lower. Leave out the walks and the 45 minutes after each, and the other windows average 42 ms, which puts the exam windows 34% lower. The walks lower RMSSD by moving the body, so they hid part of the exam’s effect. The whole-day correlation of heart rate and RMSSD, r = −0.89, blends walking and exams into one number.

RMSSD in 5-minute windows (ms) Each event Average

204060

Heart rate (bpm)

6080100

30 min before Marker 90 min after

30 minutes before

39 ms

First hour after

28 ms

Change

−29%

Locked to Exam start and averaged over 3 events: RMSSD was 39 ms in the 30 minutes before and 28 ms in the first hour after, 29% lower. The fall begins before the marker, as the student waits, and lasts as long as the exam.

The buttons work when JavaScript is on.

Two line charts of three days side by side: RMSSD in milliseconds on top, heart rate in beats a minute below, one point every 5 minutes. Shaded bands mark three exams of 90 minutes and six walks of 45 minutes; dots mark six cups of coffee. A row of buttons switches to an event-locked view, which lines up the windows from 30 minutes before to 90 minutes after each chosen marker, draws each event as a thin line and their average as a thick line, with the level of the 30 minutes before as a dashed line. Readouts and a sentence give the numbers for the view shown.

The widget makes three points that carry over to real data:

• A whole-day comparison mixes everything else in. Comparing exam windows with all other windows puts the walks, which lower RMSSD by moving the body, into the comparison. In the model the gap grows once walks are left out.
• A whole-day correlation is mostly movement. Over all windows, heart rate and RMSSD correlate strongly, because walking moves both. The number blends walking and exams and says little about either.
• Different events have different shapes. The walk gives a deep, short drop; the exam a smaller one that starts before the marker and lasts as long as the exam; coffee, which the model gives no effect, stays flat, while single coffee events still wander by a few milliseconds. That wander is why one event is not enough.

In a real study:

• Synchronise the clocks of the ECG, the accelerometer and the phone before recording, and check them at the end.
• Log markers in real time with a button or an app, and reconcile them with the diary afterwards. A marker added later from memory is only as precise as the memory.
• Fix the locked span and the baseline before analysis, for example 30 minutes before to 90 minutes after the marker, with the 30 minutes before as the baseline.
• Flag movement. Use the accelerometer or heart rate to mark windows with physical activity, as Verkuil and colleagues (2016) did, and analyse them separately or adjust for them.
• Keep the single events visible. An average can hide one event that went the other way, or one walk that fell inside the span.
• Average within each person first, then across people, or let a mixed-effects model do both.

Analysis

Within-person and between-person effects

Repeated measurements carry two kinds of information: how people differ from each other on average (between-person) and how each person changes from one occasion to the next (within-person). The two can point in opposite directions.

RMSSD (ms)

Minutes moving in the hour

• Person A
• Person B
• Person C
• Between people: slopes up
• Within each person: slopes down

A scatter plot of RMSSD in milliseconds against minutes moving in the hour. Three clusters of six dots, one per person: person A at low activity and low RMSSD, person B in the middle, person C at high activity and high RMSSD. A dashed line through the three cluster centres rises from left to right. Within each cluster, a solid line in the person's colour falls from left to right.

The example is invented. In real data, a trait that differs between people, such as fitness, can push one way while momentary activity within a person pushes the other. Curran and Bauer (2011) show how longitudinal data allow the two effects to be separated, and note that many studies do not do it. Grossman and Taylor (2007) make the HRV version of the point: within one person, RSA and vagal control are often closely related, while between people the link may be modest.

The usual fix is person-mean centring: split each repeated predictor into the person’s own average, which carries the between-person effect, and each occasion’s difference from that average, which carries the within-person effect. Enter both in the model.

Repeated measures and mixed-effects models

A mixed-effects model, also called a multilevel model, is a regression that gives each person their own intercept and, if needed, their own slopes, around the group’s average. It uses every window, copes with people who have different numbers of windows or missing ones, and estimates within-person and between-person effects together (Curran and Bauer, 2011). A repeated-measures ANOVA is a special case that needs complete, balanced data. For event-locked data, the windows are nested within events, and events within people.

Log-transform RMSSD

RMSSD is skewed to the right, and Laborde and colleagues (2017) recommend a natural-log transformation before analysis. A difference in lnRMSSD is close to a proportional change: a difference of 0.10 is roughly a 10% change.

Baseline-normalised change

Reactivity and recovery can be reported as absolute or percentage change from baseline (Laborde and colleagues, 2017). On the log scale, the difference from baseline is already a proportional change. For comparing groups, such as an intervention against a control, Vickers and Altman (2001) recommend analysis of covariance as the general approach: the follow-up value as the outcome, with the baseline as a covariate, rather than a simple change score.

Heart rate and breathing

Report heart rate next to every HRV result. De Geus and colleagues (2019) close their review with provisional reporting guidance that recognises how closely the mean and the variability of the gap between beats are tied. If you adjust HRV for heart rate, report the result with and without the adjustment. Record breathing rate in each window and say whether it stayed inside the HF band.

Multiple comparisons

HRV software reports dozens of measures, and a study has several windows and phases. Test them all and something will come up by chance. Name one primary outcome in advance, such as lnRMSSD from baseline to task. For the secondary outcomes, control the false discovery rate, the expected share of false positives among the results called significant (Benjamini and Hochberg, 1995), and label any analysis chosen after seeing the data as exploratory.

Sample size and effect sizes

Base the sample size on a power analysis with an effect size from pilot data or from similar published designs. Lakens (2013) explains how effect sizes differ between within-subject and between-subject designs: in a within-subject design the effect size can include the correlation between the repeated measures, and he argues that this suits questions about change within people. For a feel of how often an effect of a given size stands out from day-to-day noise, try the experiment in find out what moves your HRV.

Pre-registration

Pre-registration means writing down the research questions and the analysis plan before observing the outcomes, so that tested predictions can be told apart from patterns found afterwards (Nosek and colleagues, 2018). For an HRV study, register the primary measure, the windows, the artifact artifact 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 rules and the exclusion threshold, the covariates and the model. Clinical trials are registered on a registry such as ClinicalTrials.gov; the examples below show what HRV outcomes look like there.

What published studies found

Exam stress

The two exam studies that compared the exam with a calmer time agree that the heart beat faster before the exam. Lucini and colleagues (2002) found a shorter gap between beats, higher blood pressure, a higher normalised low-frequency measure and higher cortisol before the exam than 3 months later. Dimitriev and colleagues (2016) found SD1 about 29% lower and HF power about 45% lower just before the exam than at rest, with heart rate about 11 beats a minute higher. Hammoud and colleagues (2018) had no calmer day to compare with, but found SDNN highest in the hour after the exam. Their data also show how the setting muddies the picture: heart rate was lowest during the exam itself, and the normalised HF measure was highest then. The samples were students at single universities, the measures, postures and windows differed, and none shared a protocol, so the size of the effect is not settled.

Lab stressors

The TSST raises heart rate and cortisol reliably (Kirschbaum and colleagues, 1993), and in Hjortskov and colleagues’ study (2004) HF power was lower during computer work with added mental stressors than during the same work without them. A meta-analysis of 12 papers with 758 healthy volunteers found HF power lower and the LF/HF ratio higher during acute mental stress than at rest (Castaldo and colleagues, 2015). The authors read this as sympathetic activation and parasympathetic withdrawal; the HF result fits vagal withdrawal, while the LF/HF part rests on a ratio that does not measure sympathetic activity (Billman, 2013).

Reviews and meta-analyses on stress and HRV

Kim and colleagues (2018) reviewed 37 publications from 2007 to 2017 that used HRV as an objective measure of psychological stress and found that HRV changed with stress in most of them, most often as lower parasympathetic activity. Despite its title, the review summarises the studies one by one and does not pool their effect sizes, so it gives no overall figure. Together with Castaldo and colleagues (2015), the evidence supports HRV as a measure of the body’s response to stress in groups. It does not support reading a single person’s number as a stress level.

Reporting guidelines

• The Task Force (1996) set the definitions, frequency bands and recording lengths that most later work uses.
• Laborde and colleagues (2017) give recommendations for planning, analysing and reporting HRV experiments in psychophysiology; much of this page follows them.
• GRAPH (Quintana and colleagues, 2016) is a reporting checklist in four domains: who took part, how the beats were collected, how the data were prepared, and how HRV was calculated.
• Quigley and colleagues (2024) set out publication guidelines for heart rate and HRV studies in psychophysiology, starting with the physiology and the foundations of measurement.

The lesson how to read an HRV study shows how to check a paper against GRAPH.

Registered trials with HRV as an outcome

Registrations describe what a study plans to measure, not what it found. Four entries on ClinicalTrials.gov show how stress and exam studies set out HRV outcomes:

• NCT06138457, the ECOSTRESS study (University Hospital, Clermont-Ferrand): an observational study of medical students that compares HRV over 24 hours on the day of an objective structured clinical examination (OSCE) with a traditional written exam and a control day. HRV is the primary outcome; 100 students are planned and it is recruiting.
• NCT05380076 (Claude Bernard University): 490 medical students were assigned to 6 minutes of mindfulness, mobilising inner resources, cardiac biofeedback or a control video before an OSCE. The primary outcome is HRV from an ear pulse sensor during those 6 minutes, which makes it a pulse-based measure. It is completed.
• NCT02708979, StrEat (University of Copenhagen): students measured one week before an exam and again four to five weeks away from exams, with HRV from a one-minute breathing test as a secondary outcome. Its status is unknown.
• NCT05570721 (University of North Carolina at Chapel Hill): 41 midlife women did the TSST, with RMSSD at rest before the task and at 0, 10 and 20 minutes after it as secondary outcomes. It is completed.

A checklist

Before data collection

• Write the hypothesis in vagal terms, with one primary measure (usually lnRMSSD), its windows and the comparison.
• Fix the artifact rules, the exclusion threshold, the covariates and the model, and pre-register them.
• Choose an ECG sensor for tasks and daily life; if you must use PPG, check it against an ECG in your setting.
• Set the session time and posture, and plan a control session or a second baseline.
• Give participants the instructions on sleep, exercise, food, caffeine and alcohol, and plan to record medication and menstrual cycle phase.
• Work out the sample size from a within-person effect size.

During data collection

• Let each person settle for at least 5 minutes, then record a 5-minute baseline in the same posture.
• Record breathing and let people breathe at their own pace.
• Synchronise the clocks and log markers in real time; note anything that breaks the protocol, such as a coffee, a walk or a phone call.
• Watch the signal and note gaps or noisy stretches as they happen.

After data collection

• Inspect every recording, correct artifacts by the pre-registered rule, report the share corrected and apply the exclusion threshold.
• Compute the measures in windows of equal length and log-transform RMSSD.
• Fit a mixed-effects model with within-person and between-person terms; report heart rate and breathing next to HRV.
• Control the false discovery rate for secondary outcomes and label exploratory analyses.
• Report against GRAPH and Laborde and colleagues (2017), and share the beat data and code where consent allows.

Using hrv.expert tools

• The HRV Analyzer reads many recordings at once and writes one CSV of every measure. It corrects artifacts with the site’s rules or the Kubios threshold levels, reports the share of corrected beats in each recording, and writes a methods paragraph with references for the paper. Recordings are analysed in the browser and nothing is uploaded. The methods page lists every algorithm, the paper it follows and how its results were checked.
• The RR converter turns Garmin FIT, Apple Health, Polar or Movesense exports into plain RR-interval files, CSV or Kubios text, with a quality report for each recording.