.png)
Heart rate variability is one of the most clinically meaningful signals available from consumer wearables. We have made several posts on this topic, but now it is time to summarise it all in one comprehensive guide.
HRV correlates with recovery, cardiovascular health, autonomic nervous system function, stress, and sleep quality. It appears in clinical research, insurance risk models, and digital health engagement products. It is also one of the most frequently misread metrics in multi-device health platforms, because SDNN, RMSSD, and pNN50 are not the same measurement, and treating them as interchangeable produces numbers that are technically present but analytically meaningless.
Today we deep dive into what each metric measures, how they differ, which devices use which, and why normalization across devices is a real infrastructure problem rather than a minor formatting issue.
The heart does not beat with machine-like regularity. Even at rest, the interval between successive heartbeats varies slightly from one beat to the next. This variation is heart rate variability. It is a normal sign of irregularity in the clinical sense, but it is also a reflection of the autonomic nervous system continuously modulating cardiac output in response to breathing, stress, posture, thermoregulation, and dozens of other inputs.
Higher HRV generally indicates a more responsive autonomic system, better recovery capacity, and lower physiological stress. Lower HRV is associated with accumulated fatigue, cardiovascular strain, poor sleep, and in clinical contexts, with conditions including heart failure, diabetes, and depression.
The raw material for all HRV calculations is the same: a sequence of R-R intervals, the time in milliseconds between successive heartbeats. SDNN, RMSSD, and pNN50 are different mathematical operations applied to that same sequence. They each capture something slightly different about the distribution and variability of those intervals, which is why they are not substitutes for one another.
RMSSD stands for Root Mean Square of Successive Differences. It is calculated by taking the difference between each consecutive pair of R-R intervals, squaring those differences, averaging them, and taking the square root.
What RMSSD captures is beat-to-beat variability driven primarily by the parasympathetic nervous system, specifically vagal tone, the branch of the autonomic nervous system responsible for rest-and-recovery activity. Because it focuses on successive differences rather than overall variability, RMSSD is relatively insensitive to slow oscillations in the data and responds quickly to changes in recovery state.
This makes it a perfect choice for short measurement windows. A reliable RMSSD reading can be derived from a few minutes of clean heart rate data, which is why it works well for overnight wearable measurement, where devices typically have three to eight hours of continuous signal to work with.
RMSSD is the metric used in most clinical HRV research published in the last two decades, and it is the metric reported by WHOOP, Oura, Polar, and Garmin (overnight). When a wearable app talks about "your HRV," it is almost always reporting RMSSD.
SDNN stands for Standard Deviation of all Normal-to-Normal intervals. It is calculated as the standard deviation of the full sequence of R-R intervals over a measurement window.
Where RMSSD reflects parasympathetic activity and short-term fluctuation, SDNN captures total autonomic variability, including both sympathetic and parasympathetic contributions. It is sensitive to slow oscillations in the data, including influences from breathing patterns, blood pressure regulation, and thermoregulation, all of which unfold over longer time scales than beat-to-beat parasympathetic modulation.
This has two practical implications:
Apple Watch reports SDNN as its primary HRV metric. This is a deliberate choice, as Apple uses a background measurement approach that takes brief readings at irregular intervals, and SDNN is more robust to that measurement architecture than RMSSD would be.
pNN50 stands for the percentage of successive R-R interval differences that exceed 50 milliseconds. It is a count-based metric: take all the adjacent interval differences in a recording, count how many are larger than 50ms, and express that as a percentage of the total.
Like RMSSD, pNN50 reflects parasympathetic activity and short-term variability. The two metrics are highly correlated, and in most contexts they carry similar information. pNN50 was widely used in earlier HRV research, partly because it is straightforward to calculate and interpret without specialized software.
It appears less frequently in consumer wearables today, partly because RMSSD has become the preferred metric in clinical literature, and partly because the 50ms threshold is a fixed value that works better for some populations than others. For athletes with low resting heart rates, beat-to-beat differences above 50ms are common and pNN50 may be less discriminating. For older adults with higher resting rates and shorter intervals, the 50ms threshold may capture a different proportion of meaningful variability.
pNN50 still appears in some ECG-based monitoring tools, older clinical software, and academic literature. If your platform ingests HRV data from medical-grade devices or research datasets, it may be a metric you need to handle.
RMSSD and pNN50 are closely correlated and can often be treated as reflecting the same underlying physiology, though RMSSD is more precise. SDNN and RMSSD cannot be directly compared because they measure genuinely different things. A user whose Apple Watch reports 60ms SDNN and whose WHOOP reports 42ms RMSSD is not experiencing a discrepancy between devices. They are receiving two different calculations from the same underlying physiology.
For a platform connecting wearable data from multiple device manufacturers, HRV normalization is not optional. Without it, the platform is storing numbers labeled "HRV" that are not comparable across users, not trackable across time for a user who changes devices, and not usable as inputs to AI models without producing systematic bias.
The specific failure modes include:
Solving this requires knowing which metric each device reports, understanding the measurement window and methodology behind each device's implementation, and either selecting a common output metric or making the metric source explicit in every data record. It is not a problem that resolves itself through data volume.
Thryve's wearable API connects 500+ wearables and health data sources, normalizing 250+ health metrics, including HRV, across devices with different sensor methodologies, measurement approaches, and reporting conventions.
For HRV, this means the RMSSD-versus-SDNN distinction is resolved at the infrastructure layer. Partner applications and AI models built on Thryve data receive a consistent, labeled HRV stream regardless of which device a given user is wearing. The metric source is documented, the normalization methodology is transparent, and the data is structured to support both longitudinal individual tracking and cross-user population analytics without the systematic biases that come from mixing unnormalized HRV values.
For teams building insurance analytics, digital health engagement products, or clinical research platforms on wearable data from a mixed-device user base, this is the difference between a dataset that answers questions and one that generates confidence intervals around noise.
If your platform needs to work with HRV data across a mixed-device population and you want to understand how normalization works in practice, book a demo with Thryve.
Paul Burggraf, Mitgründer und Chief Science Officer bei Thryve, ist der Kopf hinter sämtlichen Gesundheitsanalysen bei Thryve und treibt unsere Forschungspartnerschaften mit der Bundesregierung und führenden Einrichtungen des Gesundheitswesens voran. Als Wirtschaftsingenieur und späterer Strategieberater entwickelte er vor Thryve mithilfe komplexer Systemdynamik grundlegende Prognosemodelle für milliardenschwere Investitionen großer Versorgungsunternehmen. Neben seiner Arbeit mit Modellanalysen und analytischer Forschung im Bereich Gesundheitssensorik ist er Gastdozent an der Zürcher Hochschule für Angewandte Wissenschaften im Life-Sciences-Master „Modelling of Complex Systems“.
.png)
.png)