This page summarizes the full specification, which is published alongside it with every weight, every curve, every limitation and the external review that changed it: Vitality Tracker Sleep Scoring Methodology, Open Specification v2.0.1, public edition (PDF), doi:10.5281/zenodo.23223863.
What it is for, and what it is not
The Sleep Health Index describes your sleep pattern against the published evidence and against your own history, for general wellness and for discussion with a health professional. It is not a medical device and it is not diagnostic: it cannot rule out sleep apnea and cannot replace a sleep study. It has not been validated against health outcomes, because no sleep score has. It does not predict how you will perform today, a claim common in this category and supported by nothing. And where a continuity measure comes from a device without a published calibration, it is not comparable with other people's scores or across devices.
Seven rules the score follows
- Weight by evidence and effect size, both disclosed. A component earns weight from the strength of the evidence and the size of the association, never from availability, visual appeal or what users expect to see. Regularity (strong observational evidence) outweighs duration (causal evidence) because its observed effect is larger and it is more modifiable; that tension is stated rather than hidden.
- The habitual pattern is the signal; a single night is mostly noise. Almost all outcome evidence concerns patterns over weeks, so the flagship number is a 28-night index and the nightly score is labeled low-confidence.
- Never score what you cannot measure. If the device on your finger or wrist cannot measure a quantity to a useful tolerance, it does not enter the score.
- Score against the right reference system. A wearable estimate is not polysomnography. It is calibrated against published bias figures or scored against your own baseline, never against a laboratory norm it was not measured with.
- Age-band only where the evidence shows an age effect. Sleep continuity changes with age and is age-banded; sleep-stage proportions do not change with age and are not.
- Never rescore history silently. Changes are versioned and published with their reasons, and past scores are recomputed with both values kept.
- Publish the criticism, not just the method. The specification carries its own external review, including the points we conceded and the three we did not.
The components and their weights
| Component | Weight | Evidence grade | What it is, and why it has this weight |
|---|---|---|---|
| Regularity | 35% | B, strong observational | The Sleep Regularity Index over 28 nights: the probability of being in the same state, asleep or awake, at any two moments 24 hours apart. In 60,977 UK Biobank adults, the most regular fifth had 30 percent lower all-cause mortality than the least regular, fully adjusted, and adding duration to the model did not improve it. It is also the sleep behavior most under your control. |
| Duration | 30% | A, causal | Hours asleep, age-banded (7 to 9 hours for adults, 7 to 8 over 65). The only component with Mendelian-randomization support for a hard outcome: short sleep raises the risk of myocardial infarction. Long sleep is not penalized, because its association with harm is most likely reverse causation; it raises a prompt to talk to your doctor instead. |
| Continuity | 15% | E, expert consensus | Sleep efficiency, onset latency and wake after sleep onset, in equal parts, scored against age-banded polysomnography norms, and only after the device's known bias has been applied or, failing that, against your own baseline. |
| Circadian alignment | 10% | C, weak or contested | Alignment between your sleep timing and your own chronotype, scored only when there is something to align to: a completed chronotype questionnaire first, your own habitual timing second, and suppressed when neither exists. Social jet lag is displayed, not scored. |
| Restoration | 10% | B | Morning sleep satisfaction (60%) and daytime alertness (40%) from validated instruments. Subjective sleep correlates poorly with measured sleep, which is exactly why it is not redundant. |
| Sleep stages | 0% | D, descriptive only | Displayed as a low-confidence estimate, excluded from the score. See below. |
The raw index is the weighted sum of the five scored components. A continuous regularity multiplier then applies, reaching full strength only when low regularity persists across two consecutive 28-night windows, so that one disrupted month, a newborn, a holiday, a bereavement, is treated as an episode, not a verdict. An earlier version capped the index at 50 below a fixed regularity threshold; the reviewer pointed out that a quintile boundary from one cohort is not a biological discontinuity, and the cliff was replaced with a curve.
Why sleep stages are not scored
This is the most visible difference between Vitality Tracker and every competing product. There is no target to hit: the "20 to 25 percent deep sleep" figure that saturates the category is a population mean reified into a goal, no trial has assigned anyone a stage percentage and measured an outcome, and the one careful study of slow-wave activity and cognition found decline at both low and high levels. The instrument cannot measure it: consumer devices classify four stages at roughly 50 to 65 percent epoch accuracy, with per-stage errors of 30 to 60 minutes a night. And even a perfect instrument would not help much at nightly resolution: with laboratory-grade EEG, a healthy person's deep-sleep share varies from night to night by about a fifth of its mean, and sleep latencies and wake time by half. The hypnogram stays in the app and the data stays exportable; what changed is that the number does not move the score, and the app tells you why. We know some people will read this as a missing feature. We would rather have a missing dial than a specification that contradicts itself in public.
Where your data comes from matters
No value enters the score without a provenance record: the metric, the device and model, the source class (laboratory polysomnography, research actigraphy, consumer wearable, sleep diary, manual entry), the firmware or algorithm version where it can be discovered, the reference system it may legitimately be scored against, and a measurement confidence. Consumer wearables systematically overestimate sleep efficiency by 2 to 10 points and underestimate wake after sleep onset by 12 to 48 minutes, and the size of the error differs by manufacturer and firmware. One tested device reported four minutes of wake against a laboratory figure of 42. Scored naively against an age norm, that user would have been handed a perfect continuity score for owning the wrong device. So continuity follows three tiers. A device with a published polysomnography validation has its bias applied before scoring. A device without one, or with an unknown algorithm version, is scored as a within-device trend against your own 90-day baseline, with the norms shown as context only. A sleep diary is scored against diary-appropriate anchors, and for onset latency the diary is arguably the better instrument.
When data is missing
A missing component is never silently replaced by the others. The index is computed over the components that exist, and the coverage, the share of the weight that was actually measured, is shown beside the score every time, together with the name of what is missing. An index of 84 with the restoration questions answered is not the same thing as 84 without them, and the display says so. The top band, Optimal, requires at least 90 percent coverage: you cannot reach it by not answering the questions.
What you see
- The Sleep Health Index, with its band (Optimal 85 to 100, Good 70 to 84, Fair 55 to 69, Needs attention 40 to 54, below 40 a suggestion to talk to your doctor), its coverage and its confidence.
- Last night's score, labeled low-confidence.
- What was measured confidently: the source device and the measurement confidence of each input.
- What cannot be measured confidently: the estimated sleep stages, marked as excluded from the score.
- The validated questionnaires clinicians use (STOP-Bang, the Epworth Sleepiness Scale, the Insomnia Severity Index, the Consensus Sleep Diary), which you can complete in the app and take to your doctor. Their results are yours to share; the app does not diagnose from them.
Where it comes from
The full reference list is in the specification. The weights rest chiefly on Windred et al., Sleep 2024 (regularity and mortality); Daghlas et al., JACC 2019 (short sleep and myocardial infarction, Mendelian randomization); Phillips et al., Scientific Reports 2017 (the Sleep Regularity Index); Boulos et al., Lancet Respiratory Medicine 2019 (polysomnography norms by age); Ohayon et al., Sleep Health 2017 (the consensus continuity indicators); Grandner et al., Sleep and Biological Rhythms 2006 (subjective versus measured sleep); and the device-validation literature (Chinoy 2021, de Zambotti 2019, Schyvens 2025).
Limitations
- The composite has not been validated against health outcomes, and no sleep score has. Its components are weighted by evidence; the whole is an untested construct, and we will not claim otherwise.
- The 80/100 anchoring of the age-banded continuity curves (the age-band mean scores 80, the young-adult reference 100) is a design convention, not an empirical result. Typical for your age does not mean optimal, and does not exclude a sleep problem.
- Device bias figures are cohort-specific, wider in older adults, and valid for a firmware snapshot rather than for the device you own today. A device moves to the uncalibrated tier the moment its firmware changes without a discoverable version. As of October 2026 no polysomnography validation of sleep staging has been published for Oura Ring 4 or Ring 5, so both sit on that tier.
- Several operational rules (naps and multiple sleep episodes, cross-midnight and travel nights, rounding, time-zone handling, record de-duplication) are listed in the specification as a completeness backlog for v2.1.0, which will ship with a reference implementation so that the specification and the code cannot drift apart.
- If the index turns out to have no better test-retest reliability than its best single component, we will say so and reconsider the composite. No competitor has published that statistic; we will.
Version
Specification v2.0.1, the public edition of 8 October 2026, which makes editorial changes to v2.0.0 of 13 July 2026 and no change to any output. This note: version 1.0, 8 October 2026. The specification is deposited on Zenodo: this version doi:10.5281/zenodo.23223863; all versions doi:10.5281/zenodo.23223862, which always resolves to the latest. Version 1.1.0 was sent for external review by an independent data scientist; the reviewer found a self-contradiction, a measurement-chain error, and a component whose inclusion broke the first design principle. Version 2.0.0 removed sleep stages from the score, replaced the hard regularity gate with a continuous multiplier, introduced measurement provenance, abolished the redistribution of missing weight, and corrected two overreaching causal claims. The three recommendations we declined, and why, are in the changelog. A change to any weight, curve, age band or threshold is a minor version with a published rationale; adding or removing a component is a major version with 30 days' notice. Historical scores are recomputed under the new version and both values are kept.
Cite and license
Cite as: Novak C. How Vitality Tracker scores sleep. Vitality Tracker Methods, method note version 1.0 for Sleep Health Index specification v2.0.1, 8 October 2026. vitalitytracker.app/science/methods/sleep-score. The full specification: Novak C. Vitality Tracker Sleep Scoring Methodology, Open Specification v2.0.1, public edition. Novak Consulting Ltd, 8 October 2026. doi:10.5281/zenodo.23223863 (all versions: doi:10.5281/zenodo.23223862).
Text © Novak Consulting Ltd, licensed under CC BY 4.0: reuse with attribution. The weights and curves may be copied; the scoring code is not covered by this license.