Methodology
How each number in the app is computed: what it reads from Apple Health, what it compares you with, and how it decides whether a link is worth showing. Written for readers who want the working, not just the result.
1. Principles
- You are compared with yourself. Every score is set against your own record — your last 30 days or nights, your usual at this hour, your usual for this weekday. Population tables appear only for orientation.
- Medians and ranks rather than means. Wearable readings carry outliers; a median, a rank or a robust slope is not moved by one wild night.
- Associations, not causes. The Stats tab reports what coincides in your data. It removes slow drifts first, so two things that merely rose together over months cannot pose as a day-to-day link, but it cannot prove a cause. An experiment can (see Test it on yourself).
- Every question asked counts. When many factors are tested against many outcomes, some will look strong by luck. The app screens all of them together before showing any (false discovery rate, q = 0.10).
- It says when it cannot tell. "Too few days yet", "could be chance" and "seen, not retained" are answers in their own right.
- Scales are declared as scales. Where a mapping is a readability choice rather than a calibrated physiological law — the Recovery curve, the sleep weights, the energy reserve's rates — the app and this page say so.
2. RMSSD and SDNN
SDNN is read as Apple Health stores it.
RMSSD comes from one of two sources:
- Computed by the app from the beat-to-beat series the Apple Watch writes to Apple Health. Intervals shorter than 0.3 s or longer than 1.5 s (above 200 or below 40 beats a minute) are rejected, and only truly adjacent intervals are compared — a gap or a rejected beat breaks the chain. RMSSD is the square root of the mean of the squared successive differences, in milliseconds, one value per series.
- Apple's own RMSSD, where the watch and the system provide it (iOS 27 / watchOS 27 and recent watches). Apple's values are used inside the time spans they cover; the app's own values remain everywhere else.
Older watches write beat-to-beat series often only when AFib History is turned on; the app says so when it sees too few of them.
3. Recovery score
- Reference: your RMSSD over the 30 days before the reading (a reading is never part of its own reference). On the watch, its own store of about a week.
- Log scale: RMSSD is right-skewed, so the app works on its logarithm.
- Outliers: values outside the quartiles widened by 1.5 times their spread are dropped (unless that would drop more than half).
- Same hour: HRV follows the clock, so each reading is compared with readings within an hour of the same time of day (15 at least), falling back to all hours when there are too few.
- Score: the distance from your usual in standard deviations is mapped onto 0–100, 50 being your usual. The mapping is piecewise linear and reaches 0 or 100 only around three standard deviations — a legibility scale, not a calibrated one.
| Zone | Score | Roughly |
|---|---|---|
| Peak recovery | 85–100 | +1.7 SD and above |
| Good recovery | 60–84 | +0.5 to +1.7 SD |
| Baseline | 40–59 | −0.6 to +0.5 SD |
| Light fatigue | 20–39 | −1.7 to −0.6 SD |
| Deep fatigue | 0–19 | below −1.7 SD |
The number shown is cut down to a whole point, never rounded up, so the figure and its zone always agree — on the phone, the watch and the complications alike.
4. Sleep score
Three components, weighted duration 35%, continuity 25%, restfulness 40%.
- Duration: 100 between 7 and 9 hours asleep (8 hours from 65). Below 7 hours it falls in a straight line to 0 at 3 hours; above the band it loses 15 points an hour, never below 60.
- Continuity: 100 minus 8 points for each awakening longer than five minutes, counted per seven hours. Minutes awake are not charged twice — they already shorten the time asleep. Fragmented sleep weighs on mood and slow-wave sleep beyond its lost minutes (Bonnet; Finan et al. 2015).
- Restfulness, from four parts, each ranked against your own nights:
- the night's HRV while asleep against its last 30 nights (35%);
- the night's low heart rate — the 10th percentile while asleep (25%);
- how activated you were going to bed — stress over the first third of the night (25%);
- the night's load — stress over the whole night (15%).
Stages are shown, not scored. Wearable staging is the least reliable part of the record, so deep, core and REM sleep appear as a share of the night beside the score.
Nights and naps: the night is the day's main sleep; other sessions are naps. A record under three hours is treated as incomplete. Labels: Poor under 50, Fair 50–69, Good 70–84, Excellent 85+.
The weights are a product calibration, checked by a lab ("Is the score built right?") against the next day's HRV. Your own 0–10 rating of a night never changes its score; another lab sets the two side by side.
5. Energy reserve
A reserve on a 0–100 scale, run forward from your stress curve over about 60 days, in the spirit of Firstbeat's "body resources".
- Asleep, the night recharges it: a night closes a share of the gap to 100 that grows with its sleep score, and the climb follows the stages (deep sleep fastest, REM slowest).
- Awake, it declines slowly, faster when stress runs above your calm daytime level; calm stretches earn a small credit, capped so that sitting still cannot refill a day.
- Watch off: only the slow baseline decline is charged — no stress is invented for hours that were not measured.
- Workouts count here (they are left out of the stress chart, not out of the reserve).
The rates are a product calibration. The energy you rate yourself is drawn on the chart and compared in a lab ("Does the curve match how you feel?"), but it never moves the curve: a rating carries mood, and at rest a feeling of energy is a vague measure.
6. Strain score
How far last night's body signals sit from your usual, read as one number. Five signals, each turned into a percentile so that higher always means more strain:
- night HRV, against its last 30 nights (lower is more strain);
- night resting heart rate, against your six months (higher is more strain);
- wrist temperature, against its last 30 nights, pulled toward neutral when the gap is within your own night-to-night spread;
- breathing rate while asleep, against your six months;
- blood oxygen while asleep, against its last 30 nights (lower is more strain).
The score is the mean of the available percentiles — at least three of five, each with eight nights or more. With fewer than five, the score is drawn toward 50 by an amount that depends on how correlated your signals are, so a partial night cannot read as extreme. The signals weigh the same, by design. Zones: Low under 35, Typical 35–54, Elevated 55–74, High from 75 (about one night in ten).
7. Stress score
For each heart-rate reading: 60% heart-rate part and 40% HRV part, scaled to 0–100. If one is missing, the other carries the score alone.
- Heart-rate part: how far your heart rate sits above your resting rate, as a share of a quarter of your heart-rate reserve (Karvonen) — resting rate from the median of the previous 7 days, maximum from 220 − age or your own highest recorded beats if higher.
- HRV part: where the RMSSD at that moment ranks in your own record (lower ranks are more stress), outliers removed; age-based values stand in until there are 50 readings.
The curve is smoothed (a short rolling median, then a 15-minute exponential smoothing). Workouts, and five minutes after them, are left out. Zones: Calm under 25, Balanced 25–49, Stressed 50–74, High 75+. Steps do not enter the score; a lab reads a version of it with the heart rate explained by your steps removed.
8. Stats: what moves your numbers
Top influences
- Six months of days. Each factor (bedtime, steps, daylight, workouts, caffeine, alcohol, outdoor temperature, mindfulness…) is set against each outcome (RMSSD, sleep score, stress, resting heart rate, breathing, sleep duration and stages, blood oxygen, blood pressure, felt energy). A factor of the day is read against the night that follows.
- Daily anomalies: both series have their slow drift removed first (a robust Theil–Sen line), so a season or a habit change cannot pass for a day-to-day link.
- Effect size: for a continuous factor, the robust slope times the factor's typical range, as a % of your usual (or in points for a score); for a yes/no factor, the median on days with it against days without, five days a side at least. Below three tenths of the outcome's own day-to-day spread, an effect is called "no clear effect" however significant.
- Direction must hold: a rank correlation of at least 0.1 pointing the same way, and the sign must survive 90% of 300 bootstrap resamples.
- The chance screen: each link gets a p-value from the calendar itself — the factor's days shifted or rotated against the outcome's, which keeps both series' own rhythms intact. Related gauges are combined (Simes), then a discrete Benjamini–Hochberg procedure at q = 0.10 runs over every question asked of an outcome.
- Confounding check: after the screen, a continuous link that strongly tracks a factor from another family is re-read with that factor held constant; it is demoted if it vanishes or reverses, relabelled if it weakens.
- No single direction: some factors are better close to your usual, or away from it (irregular bedtimes, for example). Those links are screened among themselves and shown with two figures: the lowest quarter of days and the highest quarter, each against the middle half.
- Seen, not retained: a link whose own test saw something but that the screen set aside keeps its figure, greyed and never ranked, so a link at the edge does not appear one day and vanish the next.
Your own log
A habit you log is read against the night that follows — night HRV, sleep score, night heart rate — each night as its gap from the median of its 30 neighbours. A blank day is no answer, never a "no". Questions: yes against no (five days a side); bigger against smaller amounts (ten days with an amount); three-day means (for water, which acts over days); and the change per extra unit (a robust slope, twenty days). All logged comparisons are screened together (calendar null keeping runs of days whole, Benjamini–Hochberg at q = 0.10, bootstrap sign at 90%). A reading close to the bar — sign stable, reached by chance no more than one time in five — is shown greyed as "seen, not retained".
Hours after a session
RMSSD in the 6 hours after a workout, or the 3 hours after a mindfulness session, against the same clock hours on days without one; a day counts once, five days at least. A figure appears when the 95% range across days excludes zero. The hours before are read the same way: when they already read like the hours after, the figure is the day's, not the session's, and is drawn without colour.
9. Long-term trends and moving together
- Drift cards (on each page): your usual — the median of the previous 30 nights or days — and how far it moved over one, three or six months. It turns coloured past half of a usual day's spread.
- Long-term trends (Stats): six measures read like their drift cards; the others compare the recent half of six months with the first. A direction is shown only after a rank-correlation test against time by calendar rotation, screened across the family.
- Moving together: weekly means over whole weeks; the week-to-week changes of two series, detrended, compared by rank correlation; p-values from shuffling blocks of three weeks (sequential stopping, Besag–Clifford), then Benjamini–Hochberg at q = 0.10 and |ρ| ≥ 0.25 to certify a link. Correlations between weekly levels are shown as description only.
10. Labs
Each lab asks one question of your record, answers it in a sentence, and shows its working. A lab never changes a score, and the coach does not quote the labs.
| Lab | Method |
|---|---|
| Where is it drifting? (HRV, resting heart rate, sleep, nights, days) | One card per page. Your usual is the median of the previous 30 nights or days; the card says how far it moved over one, three and six months, against half of how far a day usually sits from it. A measure the watch reads differently restarts its usual after the change of watch, and the step is said apart. |
| Did something change, and when? | Seven measures (RMSSD, resting heart rate, sleep score, sleep duration, bedtime, wake time, steps), as weekly medians over 26 weeks (four days a week at least). The split where the weeks before and after differ most, weighed by the weeks on each side (four at least), is tested against 999 shuffles of whole weeks — a day carries over into the next two or three, so shuffling days would overstate certainty. The 5% bar is shared across the measures; the gap must reach half a usual day's spread. Because the result can hang on which weekday a week begins, all seven ways of cutting the weeks are read: the change is dated at the middle of their dates, and its chance is the middle of theirs. A split reached by chance up to one time in ten is shown as "seen, could be chance". |
| Which stretches stood out? | Each day carries the median of the seven days ending on it. That median is "far" when it leaves its own ordinary range over six months (Tukey fences: the middle half widened by 1.5 times its width). A stretch needs enough area past the fence — a few days well past it, or more days just past it — calibrated so that ordinary ups and downs produce one about once in twenty six-month records. It is dated to the days nearer its level than your usual. Beside it, for information only: the other series that were more extreme over those days than over any other run of the same length around that time of year, with how many series any such run would show by chance. |
| Does a low stretch last? | Runs of days below the median of their previous 30, counted once per run, against the base rate. |
| Does the afternoon announce the night? | Afternoon RMSSD against the next night's score, each against its own previous 30 days; never across a change of watch. |
| Is it you, or the reference? | The Recovery score against its moving 30-day reference and against a frozen one, to separate a change in you from a change in the window. |
| Is the stress curve just your heart rate? | Six versions of the stress score side by side (heart rate only, "at rest" with steps removed, a 30-day reference, …) and what each adds. |
| Is the score built right? · Does the score match how you slept? | The sleep score and its parts against the next afternoon's HRV; your 0–10 night ratings against the score, read on each half of the record (opens at 20 rated nights). |
| Is the reserve the night's score again? | Whether the morning reserve says anything the night's score does not, read as anomalies against the surrounding 30 days. |
| Does the curve match how you feel? | Your energy ratings against the reserve with each hour's usual removed, read separately after moving and at rest. |
| Day read | How today differs from the same weekday, at the same hour of your day. Descriptive. |
| Test it on yourself | A randomised n-of-1 trial over 28 days, in blocks of two days on and two days off; a randomisation test at 5%, the result hidden until the end. |
| Fitness age · How you walk · Compared with other people | VO₂max against the FRIEND registry norms; walking measures against their own noise; your values as percentiles of published cohorts. |
11. A change of watch
A new watch, or the move to Apple's own RMSSD, can read a measure lower or higher overnight — on the first user's record, Apple's RMSSD reads nearly a quarter lower than the app's own calculation. The app keeps one record for all watches, and handles the step where it matters:
- labs and drift cards read a measure the watch moves on each side of the change, never across it, and report the step apart;
- scores built on a 30-day reference absorb the step within about a month, and a lab shows how much of a move is the reference;
- a change found close to that date in a score built from such readings says it may be the watch.
12. The on-device coach
The coach's notes are written by Apple's on-device language model (Apple Intelligence, iOS 26 and later). It is given fact lines the app has already computed and chooses and rewrites them as prose. Its output is checked before it is shown: a note that invents a number, flips a direction, adds a claim or upgrades the night's verdict is refused, and the facts are shown as written instead. Without the model, the app writes the note itself. Nothing is sent to a server.
13. Limits
- Wrist sensors are not an ECG. Readings during movement are noisier, and sleep stages from a watch are estimates.
- Many weights and rates are product choices, tested against your own data where a lab can — not clinical standards.
- Associations found in one person's record describe that record. They can point to an experiment; they do not establish a cause.
- HRV Tracker gives wellness information, not medical advice. If something worries you, see a doctor.
Sources for the physiology and the statistics are on the scientific references page.