Sleep Analysis

Sleep Analysis is a sequence of time intervals labeled with a sleep state. It can answer when sleep was recorded, how much elapsed time was labeled asleep, and how the estimated stages changed. It is not a single sleep-quality score or a diagnosis.

Each row has a start, an end and one category value. The number of rows describes how finely the night was segmented, not the number of nights slept. A quiet, continuous block may be one row; a night with many stage changes may contain dozens.

The categories describe different things, so they need different denominators. Apple's sleep-category documentation explicitly permits In Bed to overlap the detailed sleep stream.

Core includes N2 and some N1 in Apple's definition. A stage label is a device estimate, not a direct brain-wave measurement. Avoid treating a short Deep segment or an absent REM label as proof of a disorder.

The figure below shows actual intervals from example user data, with dates and device identity removed. The horizontal axis measures elapsed hours rather than wall-clock time. Long bars are sustained recorded states; short bars are brief segments.

There were no In Bed rows in the example year-long snapshot. The empty lane makes that absence visible. It means time in bed cannot be calculated from these records; it does not mean the person never spent time in bed. Apple notes that Watch Awake samples occur between sleep samples, so their total can also omit wakefulness at the beginning and end of a session.

There is no fixed row cadence. Apple describes classification windows in its sleep-stage validation paper, but consecutive classifications can become a longer stored interval. A row boundary is a recorded state transition, not an independent reading every thirty seconds.

In the researched year, 10,941 intervals covered 363 of 364 eligible reporting nights after excluding truncated boundary bins. Median individual segments were about 15 minutes for Core, 8.5 minutes for Deep, 19 minutes for REM and 2 minutes for Awake. Those figures describe this dataset's segmentation, not expected stage durations per night.

Short daytime episodes need separate attention. The analysis found 18 candidate episodes lasting under three hours; none carried detailed Core, Deep or REM stages. These were identified by time and duration, so calling every one a confirmed nap would overstate the evidence. Apple's 2025 paper describes version-dependent binary sleep/wake reporting for shorter sessions. Preserve Asleep unspecified rather than dropping it because staging is missing.

Choose your sources, retain categories 1, 3, 4 and 5, and calculate the union of their intervals. Do not add In Bed or Awake. Unioning avoids counting the same elapsed minute twice when sources or stages overlap. It does not decide which conflicting stage label is correct; conflicting detailed labels need a separate source-selection policy.

The next figure compares all-row duration totals with asleep-only totals and the asleep union. In the example records, asleep row sums and the union agreed to within tiny timestamp-boundary differences. The substantial excess came from adding Awake, not from duplicate asleep sources. Across the 92-night focus window, that mistake added a median of about 24 minutes per reporting night.

For this case, median asleep union time was about 7.6 hours in the focus window, with the middle 90% of recorded nights spanning about 5.6–8.9 hours. This is an observed example distribution, not a clinical normal range. In Bed was unavailable, so sleep efficiency and time-to-fall-asleep were not calculated.

A night crosses midnight, so grouping by sample start date can split it. This analysis used local noon-to-noon bins, labeled by their ending date, in one declared reporting timezone. Intervals crossing a bin boundary were split. Durations used UTC elapsed seconds, avoiding an extra or missing hour during daylight-saving transitions. First and last bins were excluded because the export could truncate their sleep episodes.

Noon bins may combine overnight sleep and daytime sleep. For questions about naps or sleep timing, also reconstruct episodes with an explicit gap rule; our exploratory analysis separated episodes after a gap greater than 90 minutes within a source. This is an analyst rule, not an identifier provided by HealthKit.

There were no zero-length, reversed or unknown-category rows. Statistical duration outliers were retained for review. A short recorded night might represent genuinely short sleep, incomplete capture, or a grouping boundary. Missing nights and gaps cannot establish non-wear, wakefulness or the absence of a physiological stage. Two Watch source groups and their software revisions were audited separately; apparent shifts can reflect a source or algorithm change.

Align heart rate and HRV observations to actual asleep intervals before comparing nights. In this dataset, a typical recorded night had about 107 heart-rate samples but only four HRV samples overlapping sleep. HRV is a sparse set of observations, so a nightly median does not describe every stage or every minute.

The analysis also compared respiratory-rate medians within asleep intervals, and matched workouts from the noon-to-noon window immediately before the sleep reporting bin. That workout window ends at noon before the night, so it excludes exercise later that afternoon or evening. Night-level sleep duration showed little linear association with the physiological medians in this example. This is a description of the selected records, not evidence that sleep cannot affect these metrics. Exercise, illness, timing, device behavior and recording density can confound comparisons; a workout record's absence does not establish inactivity.

Use consistent-source trends to explore sleep timing, recorded duration and estimated architecture. Keep personal observations distinct from medical guidance. Apple's classifier is based on accelerometer-derived signals and validated against polysomnography; its paper describes updates in 2024 and 2025 that can change longitudinal comparisons. The stored sleep intervals do not reveal a proprietary stage-detection rule based on an individual's HRV or respiratory samples.

If sleep feels unrefreshing, you have persistent difficulty sleeping, or someone notices breathing problems at night, discuss the symptoms with a clinician. A consumer stage chart cannot diagnose a sleep disorder or replace a clinical assessment. Sleep apnea events, where available on compatible systems, are a separate HealthKit concept rather than something inferred from these stage intervals.