How much does resting heart rate have to change before it stops being noise? The threshold from 650 nights

A single resting heart rate reading tells you almost nothing. The question is not whether the number moved, but whether it moved beyond the noise floor. With 28 nights of continuous history, the minimum threshold is 5.86%. Below that number, the movement may be normal statistical variation. This is not an estimate. It comes from 650 nights of a single body, with no interpolation and no imputation.

Why a single reading fails

Resting heart rate looks stable. It is not.

In this dataset, the coefficient of variation (CV) for resting heart rate is 7.45%. That means the metric drifts up and down with a natural dispersion equivalent to 7.45% of the central value, with no real-world intervention required. This is intrinsic system noise — not a device malfunction, not a measurement error, but the baseline volatility of a living system under continuous observation.

The median resting heart rate across these 650 nights is 44 bpm. With a CV of 7.45%, the natural swing between consecutive nights is substantial. A higher or lower reading today compared to yesterday may be noise and nothing else.

The daily autocorrelation of the metric is 0.386. Today's value has a moderate correlation with yesterday's, not a strong one. Prior context informs but does not predict reliably. An autocorrelation near zero would mean each night is statistically independent of the previous one. An autocorrelation near one would mean yesterday's reading nearly determines today's. At 0.386, yesterday's value carries partial information — enough to matter for the calculation, not enough to anchor the estimate.

What happens when you compare two isolated readings? The standard threshold for declaring a real change between two single data points — the Reference Change Value, or RCV — is 20.66%. That threshold is high because the noise in one reading versus another is large.

No one waits for a swing that wide before drawing conclusions. The practical fix is to accumulate nights and let the noise average out.

How time compresses the threshold

Building a continuous history lowers the threshold for real change. Each additional night narrows the confidence interval around the true baseline. Noise does not vanish. It averages.

Each night is another sample of the underlying true baseline. The more nights you stack — uninterrupted, with the same device — the tighter the estimate of that baseline becomes. The threshold shrinks not because the physiology grows more predictable, but because the statistical picture of what is normal for this body becomes more precise.

The table below comes directly from ruido.medido.json#resting_hr:

Parameter	Value
Nights measured	650
Median	44 bpm
CV	7.45%
Autocorrelation (ρ)	0.386
RCV (1 reading)	20.66%
Threshold at 28 nights	5.86%
Threshold at 56 nights	4.15%
Threshold at 84 nights	3.39%
Threshold at 168 nights	2.39%

At 28 nights, the threshold drops from 20.66% to 5.86%. At 56 nights it falls to 4.15%. At 84 nights, 3.39%. At 168 nights, the threshold reaches 2.39%.

The instrument becomes more precise as history accumulates. But only if that history is continuous and collected with the same device.

How the threshold was derived

The dataset contains 650 nights from a single body, recorded with the same wearable device without interruption. They are part of 18483 measurements across 31 metrics.

No data was interpolated. Nights without a valid reading were excluded, not filled in. The method applies clinical metrology principles to continuous wearable data. Clinical metrology is the discipline that governs how measurement uncertainty is calculated and communicated in health contexts — how to state not just what was measured, but how reliably, and what magnitude of change would be required to exceed the instrument's own noise floor. That standard is rarely applied to consumer wearable data. It is applied here.

The CV (7.45%), autocorrelation (ρ = 0.386), and RCV (20.66%) were computed across the full 650-night series. Window thresholds — u28, u56, u84, u168 — were derived from those base parameters, adjusted for window size.

This is not a population value. It is the value of one body over one period. That specificity is why the method is useful: it captures the noise profile of this person, not an average across many different people with different physiology.

What the threshold means in practice

Suppose you have 28 nights of clean history. Your resting heart rate drops by a percentage greater than 5.86% from your baseline. That movement crosses the statistical threshold. There is signal.

If the change is smaller than 5.86%, the data neither confirms nor refutes that something happened. That is not the same as saying nothing happened. It means the instrument, at that window size, does not have enough resolution to separate the movement from noise.

The distinction matters. Conflating "the data does not detect it" with "the data says there is no effect" is a systematic error in the wearable health space. Most consumer devices do not publish their resolution limits. They display a value — a resting heart rate of 44 bpm, for example — without indicating what change from that value would be signal versus noise. The number moves; the user assumes the movement is meaningful. The instrument has resolution limits. Publishing those limits is part of the method.

At 56 nights, resolution improves: the threshold drops to 4.15%. At 84 nights, 3.39%. At 168 nights, the instrument reaches 2.39% — the sharpest threshold available from this dataset.

The minimum entry point for a valid analysis is 28 continuous nights. Before that, the confidence interval is too wide to declare signal with rigor.

The limit

This threshold belongs to one body, measured with one device, over a bounded period.

It is not universal. A different body has its own CV. The same body at a different time may have a different CV if baseline conditions shift — sleep, training load, chronic stress. A different device introduces its own systematic bias. The thresholds from this dataset do not transfer directly to another body or another device.

The autocorrelation of 0.386 means the model assumes moderate day-to-day correlation. If your week has strong cycles — high stress during workdays, recovery on weekends — the real behavior may diverge from the model's assumption.

This dataset does not say what caused the changes it records. It identifies when a change carries enough statistical weight to be more than noise. Causation is a separate question, and this file does not answer it.

The thresholds published here are from 2026, computed on 650 nights. If the history grows, the parameters are recalculated and the thresholds may shift.

Further reading

• Mental Work — all protocols
• Related file: magnesium-for-sleep
• Open science — the dataset