Same effort, different life, different recovery.
Deep sleep is the body's scarcest, most restorative currency. It is the stage that does the heavy structural repair, and the one most easily lost. Across 431 nights, this user ran a structural deficit: a mean of about 33 minutes a night, roughly 8% of total sleep, against a 13 to 23% healthy reference.
The interesting part is what moved it. Not training. Across the year, the largest levers on deep sleep were environmental and behavioural: daylight, life stress, and a couple of small evening habits. The training calendar barely registered against them. The body was keeping score of the whole life, not just the sessions.
Same effort, different life, different recovery.
How to read this case.
As in the first case study, every finding is read in two layers. The first is the biomarker: deep sleep, REM, HRV, resting heart rate. The second is the context: daylight, life events, habits, and what went in the body the night before. The number tells you what happened. The context tells you why.
A small recovery vocabulary runs through the piece. The HRV readiness floor sits around 30ms, below which the autonomic system is clearly taxed. The resting-heart-rate recovery floor sits around 47bpm, the marker of a fully rested, fully tapered state. And REM and deep sleep respond to different stressors, which is why they sometimes move in opposite directions.
Daylight moves deep sleep.
The cleanest environmental signal in the dataset. Across 385 paired days, the sunniest quarter of days produced about 30% more deep sleep than the darkest quarter: 45.6 versus 35.2 minutes a night. The effect holds at +22% even with the trip period excluded, so it is not just an artefact of one bright stretch.
The relationship is not linear, and the data says so honestly. Day-to-day, the correlation is weak (Pearson 0.14); daylight is one input among many and the daily noise is large. The story is in the tail. The step-up concentrates above roughly two hours of daylight, and this user was below that threshold on about 82% of normal-life days.
Stress stacks, it doesn't add.
The intellectual core of the year. A natural experiment made it legible: a 100-day period living away from normal work and routine, compared against the prior 100 days of ordinary life. Training continued throughout both. Life stress did not.
The payoff is in the decoupling. Bucket the days by training load and compare HRV within each bucket: at every matched load, HRV ran 3 to 8ms higher away from normal life. And the gap widened with intensity, reaching about +8.3ms at the hardest sessions. If training load alone explained recovery, the gap would close at matched load. It does not.
Two more readings point the same way. There were zero sub-30ms readiness days during the low-stress period, against 4% in normal life. And same-day HRV immediately after a hard session was near-identical between the two periods: the hard session lands the body in the same place, but in the low-stress period that place was further from the floor to begin with.
The body recovers from training and life stress out of the same budget.
The alcohol fingerprint.
A clean dose-response signature, and a small sample. This is four nights, so it is suggestive rather than definitive, and the section keeps that caveat in view. But the fingerprint is consistent and worth reading.
The signature is selective. REM is suppressed, down to −62% on the heaviest night, while deep sleep is preserved or even slightly elevated as early-night sedation deepens it. HRV and resting heart rate carry the metabolic cost of overnight clearance. Two tiers separate cleanly: heavy and late (HRV crashing 21 to 63%, RHR up 9 to 15bpm) versus moderate at a normal bedtime, which barely registers.
The fingerprint is REM, not deep sleep. Acetaldehyde suppresses REM while early sedation can deepen N3, which is why a drinking night can look "deep" on a summary while the restorative REM work quietly goes missing.
The evening habit change that worked.
The hopeful close, and the cleanest single-variable improvement in the dataset outside the trip. In early 2026 two evening changes happened together: food capped earlier in the evening, and screens reduced after a set time. Bounded windows were compared on either side of the change, chosen to avoid race-taper confounders.
The result was a 35% reduction in awake-during-night time, about 9% more deep sleep, and higher sleep efficiency, even at a slightly shorter total duration. REM was unchanged. An extended 37-night window confirmed the direction.
What the year shows.
Set the findings against the thesis and a ranking appears. The largest swing of all is seasonal: deep sleep ran about 58% higher in summer than in winter, the combined circadian envelope that the daylight finding sits inside. Then daylight itself, then the evening habit change, with life stress acting across all of them as the autonomic modifier rather than a single lever.
| Lever | Effect on deep sleep | Type |
|---|---|---|
| Seasonality | +58% summer vs winter | Environmental envelope |
| Daylight | +30% annual / +22% normal life | Environmental |
| Evening habit change | +9% deep, −35% awake | Behavioural |
| Life stress | +3 to +8ms HRV at matched load | Autonomic modifier |
| Training load | Little independent effect on deep sleep | Reference |
Two smaller threads close the year. A magnesium-based sleep supplement showed an early, unconfirmed signal across a short window, the REM-to-deep ratio shifting from about 3.0:1 toward 2.2:1. It is 17 nights with a post-race confounder, so it reads as a signal worth watching for another clean month, not as efficacy. And the seasonality result is the envelope the daylight finding lives inside, not a separate cause.
Where recommendations appear, they belong to the data or to a named human, not to Me. The data associates more daylight, earlier evenings, and lower background stress with better deep sleep. Where a target was set, it was set by the user's coach, for instance reaching the resting-heart-rate recovery floor a set number of times in race week, an instruction that belongs to the coach and the data behind it.
Methodology.
Data sources. All data was collected via Apple Health on-device: sleep stages, HRV (rMSSD), resting heart rate, respiratory rate, wrist temperature, workouts, and time in daylight. Annotations and check-ins were entered in the Me. app. The window covers 431 nights across 14 months.
Known limitation on deep sleep. Apple Watch is understood to underestimate N3 deep sleep by roughly 15 to 25%, so the stage percentages here are conservative. The relative comparisons (sunny vs dark, before vs after) are more reliable than the absolute minutes.
Baselines and exclusions. All baselines are personal averages across the window, not population norms. A block of duplicate stage records (16 to 24 March 2025) produced impossible durations and was excluded as a data-quality measure.
Statistical methods. Pearson correlation where the relationship was plausibly linear, quartile comparison where it was not, and a matched-load comparison for the stress decoupling. Several variables were not measured, including caffeine, bedroom environment, sleep-onset latency, within-night stage timing, endocrine markers, and the precise count of awakenings; conclusions are bounded accordingly.
What this case study is not. It is not a medical assessment, and nothing here was diagnosed by a clinician. Any respiratory-rate or temperature signals are read as subclinical observations, not diagnoses. The case study demonstrates a reading framework, not a diagnostic tool. To understand what any of this means for you, take it to your GP or compile it for analysis. Not medical advice.