Beacon · Research

Every metric, traceable
to its paper.

Beacon computes five metrics. Each one comes from a published, peer-reviewed method, and each is listed below with its citation, its identifier, what the paper established, what we implemented from it, and the input the method requires. Where our implementation departs from the original, that is stated in full rather than left to be discovered.

Sleep Regularity Index

Phillips AJK, Clerx WM, O’Brien CS, et al. Irregular sleep/wake patterns are associated with poorer academic performance and delayed circadian and sleep/wake timing.

Scientific Reports 2017;7:3216

doi:10.1038/s41598-017-03171-4

What it established: A single index for how repeatable a sleep pattern is from one day to the next: the average concordance of sleep/wake state between all epoch pairs exactly 24 hours apart, rescaled so that 100 is perfectly regular and 0 is chance.
What Beacon computes: The index itself, computed over the supplied recording window using the same construction as published, with no substitution of a shorter window when the input is too short to support it. It is returned as a subject-scoped value: computed across a subject's consecutive recording days, attached to each of those nights with the same number, and never presented as a property of one night.
Input required: Epoch-level binary sleep/wake over multiple consecutive days. It is not computable from a single night: Beacon returns it as unavailable, with the reason, rather than a number derived from too little data.

Evidence

Cohort
Windred DP, et al., SLEEP 2024;47(1):zsad253. 60,977 UK Biobank participants, over 10 million hours of wrist accelerometry, mean 6.3-year follow-up, 1,859 deaths.
Mortality
Most-regular versus least-regular quintile: all-cause mortality HR 0.52 (95% CI 0.45–0.60) minimally adjusted, 0.70 (0.59–0.83) fully adjusted.
Versus duration
Sleep regularity outperformed sleep duration as a mortality predictor, and duration added no significant incremental variance once regularity was in the model.
Cardiometabolic
Lunsford-Avery et al., Sci Rep 2018;8:14158 (MESA, n=1,978, mean age 68.7): 10-year ASCVD risk partial r = −0.133, BMI r = −0.139, HbA1c partial r = −0.138.

Odds Ratio Product

Younes M, Ostrowski M, Soiferman M, et al. Odds ratio product of sleep EEG as a continuous measure of sleep state.

SLEEP 2015;38(4):641–654

doi:10.5665/sleep.4588

What it established: A continuous measure of sleep depth in place of discrete stages. EEG is split into 3-second epochs and Fourier-transformed into four bands. The bands are delta 0.3–2.3, theta 2.3–7, alpha-sigma 7–14 and beta 14–35 Hz. Each band is ranked into deciles, forming a 4-digit code with 10,000 possible patterns. Each pattern maps to its probability of occurring in wake-scored epochs, divided by 40, giving a continuous 0–2.5 scale where 0 is deepest sleep and 2.5 is fully awake.
What Beacon computes: The 3-second series and its 30-second aggregate. Ten 3-second values average to one 30-second value, as published.
Input required: Raw EEG. A hypnogram alone is insufficient; the measure is defined on the spectrum, not on stage labels.

Where our implementation departs from the paper

Two departures, both stated on every ORP value the API returns. First, the original normative lookup table is proprietary to Cerebra Health, so Beacon re-derives the decile table from the open Sleep-EDF corpus rather than reproducing the vendor table. The construction is the same, the fit is independent, and detail.table_source records which table produced the value. Second, that re-derivation widens the reference class from the published wake-only class to W+N1, and says so in detail.reference_class: the open corpus Beacon fits on does not contain enough wake to populate the sleep end of a wake-only scale, and fitted that way the N2, N3 and REM means collapse to roughly 0.15 each, losing the stage ordering the paper reports. Widening the class restores that ordering. The consequence is that Beacon ORP values are on their own scale and are not numerically interchangeable with values from the vendor table.

Sleep spindle density

Lacourse K, Delfrate J, Beaudry J, et al. A sleep spindle detection algorithm that emulates human expert spindle scoring.

Journal of Neuroscience Methods 2019;316:3–11

doi:10.1016/j.jneumeth.2018.08.014

What it established: A spindle detector called A7, designed to emulate human expert spindle scoring rather than to trip a fixed amplitude threshold.
What Beacon computes: Spindle density: spindles per minute of N2, using the A7 implementation in YASA (BSD-3-Clause).
Input required: Raw EEG plus a hypnogram, and at least 5 minutes of N2.

Slow-wave activity decay

Borbély AA. A two process model of sleep regulation.

Human Neurobiology 1982;1(3):195–204

https://pubmed.ncbi.nlm.nih.gov/7185792/

What it established: The two-process model: homeostatic sleep pressure, termed Process S, dissipates across the night alongside a circadian process.
What Beacon computes: The exponential decay of delta (0.5–4.5 Hz) power across NREM epochs, reported as the decay constant with its R-squared, so the fit can be judged rather than taken on trust.
Input required: Raw EEG plus a hypnogram, and at least 20 NREM epochs.

Sleep–wake transition dynamics

Lo CC, Chou T, Penzel T, et al. Common scale-invariant patterns of sleep-wake transitions.

PNAS 2004;101(50):17545–17548

doi:10.1073/pnas.0408242101

What it established: Wake-bout durations follow a power law, with an exponent near 1.3 that is conserved across species, while sleep-bout durations are exponential.
What Beacon computes: The power-law exponent, fitted by maximum likelihood, together with stage-transition entropy.
Input required: A hypnogram only. No raw EEG.

Limits

What we do not claim

Beacon implements published, validated algorithms on real recordings. That is a statement about method, not about regulatory status. Beacon is not a cleared medical device and is not intended for diagnosis.

Reference implementations run on PhysioNet’s Sleep-EDF Expanded corpus, whose hypnograms were scored by human technicians.

physionet.org/content/sleep-edfx/1.0.0

Demo

See these computed on a real night

Docs

API reference