The wearables are good now. Oura and Whoop hit polysomnography-level accuracy on total sleep time and sleep efficiency, and they get architectural staging into the high-80s percent agreement on most nights. What they cannot tell you is why a Tuesday was bad and a Wednesday was good. The patterns that matter for sleep optimization usually live in subjective and contextual data the device can't capture: how mentally loaded you were at bedtime, whether you woke at 3am because of a bathroom run or a cortisol spike, whether the dream you remember was the third REM cycle or the fifth.
A paper diary catches all of that. Four columns are enough.
What the wearables miss
Chinoy et al. 2021 ran consumer wearables (Oura, Whoop, Apple Watch, Garmin) against polysomnography in a controlled lab cohort ( Chinoy et al. 2020, n=8 ). Headline numbers:
- Total sleep time: within 10-15 minutes of PSG on most devices.
- Sleep efficiency: high agreement, generally within 2-3 percentage points.
- REM staging: 70-80% agreement with PSG. The 20-30% misclassification is mostly REM-as-light or light-as-REM.
- Deep sleep staging: 60-75% agreement. This is where consumer devices struggle most.
The takeaway is not "wearables are wrong." It is that for staging-specific decisions (am I getting enough deep sleep?), there's about a 20% noise floor. For total sleep and efficiency, the data is excellent. For dream recall, presleep arousal, and the texture of awakenings, the device cannot record what is happening inside your head.
Walker 2017 ("Why We Sleep") and Ohayon 2004's stage-distribution reference work both ground the architectural reference ranges that wearable apps display in graphs ( Walker 2017 Ohayon et al. 2004 ). The graphs are useful. The graphs do not tell you that the bad-deep-sleep night was preceded by 2 glasses of wine at 9pm.
The four columns that matter
The minimal viable sleep diary fits on a single page per week. Four columns, one row per night:
- Bedtime + lights-out time. Bedtime is when you got into bed. Lights-out is when you committed to sleep. The gap matters.
- Sleep latency (minutes). Estimate. You can't time it precisely, but you know whether it was 10 minutes or 60. Just write the number.
- Mid-night wake count + total awake duration estimate. How many times did you wake. Roughly how many minutes total were you awake during the night.
- Morning alertness, 1-10. Subjective rating right after waking. 1 is "I cannot function without coffee" and 10 is "I jumped out of bed."
Optional fifth and sixth columns for power users:
- Notable presleep input: alcohol, late caffeine, hard workout, screen time, conflict, deadline.
- Dream recall (yes/no, brief). Tracks REM, indirectly. Most dream recall comes from the last REM cycle of the night, which is the longest. Consistent dream recall is a soft proxy for adequate REM.
Mental load is the column that surprises people. Bedtime mental rumination is one of the strongest predictors of sleep latency in DSM-5-screened insomnia, and it is invisible to every wearable on the market.
Why 14 days is the floor
One night tells you nothing. The within-person variability in sleep latency, wake count, and morning alertness is high enough that any single night could be noise. Three nights captures most of the noise distribution but doesn't yet show patterns.
Fourteen days is the floor for these patterns to surface:
- Weeknight-weekend drift: did your bedtime drift more than 60 minutes between weekday and weekend? Social jet lag is real, and it predicts metabolic and mood outcomes.
- Day-of-week clusters: if Thursday morning is consistently a 5/10 and Friday is consistently a 7/10, you have a Wednesday-night signal worth investigating.
- Input correlations: alcohol on Wednesday → mid-night waking on Thursday is the kind of pattern that's invisible on any individual night and obvious across two weeks.
- Cycle effects (for women): half a cycle. Two weeks captures roughly half a menstrual cycle, enough to flag the late-luteal sleep disruption that's a clinical pattern.
Beyond 14 days, marginal information per additional night drops fast. The serious clinical sleep diaries used in chronic-insomnia treatment run 14 days; that's where the convention comes from.
Where to put the wearable data
The diary doesn't replace the wearable. It complements it. If you're already wearing an Oura or a Whoop, the integration is simple: add a seventh column for "wearable score" (sleep score 0-100, or whatever the device emits). Then the comparison is direct.
Two patterns to look for:
- Wearable says "good," you say "bad": the device captured normal architecture but you woke unrefreshed. Common causes: presleep arousal, fragmented light sleep that the device read as continuous, undiagnosed sleep-disordered breathing. If this is consistent, it's worth a clinical sleep evaluation.
- Wearable says "bad," you say "fine": the device under-counted your deep or REM, often because of motion noise (HRV-based devices misread restless light sleep). The subjective reading is more reliable here.
When the device and the diary disagree consistently in the same direction, the diary is the truer signal.
The clinical handoff
If sleep is bad enough that you're considering a sleep medicine consult, walking in with a 14-day diary in hand changes the consult. Cain et al. 2023 in the mental-health context found that subjective sleep diaries shifted clinical recommendations more reliably than wearable data alone, and a clinician can read a 14-day paper record in 60 seconds ( Cain et al. 2023 ).
The tracking patterns the clinician will look for: sleep efficiency under 85% (time asleep / time in bed), latency over 30 min on >50% of nights, mid-night wake duration over 30 min on >50% of nights. Those three thresholds, applied to a 14-day record, are the DSM-5 insomnia screen.
What this isn't
This is not a replacement for polysomnography for sleep-disordered breathing. If the diary surfaces consistent loud snoring reports from a partner, frequent gasping awakenings, or daytime sleep attacks, those are PSG signals, not diary signals. The diary is for the much larger population whose sleep is functional but suboptimal, and where the patterns matter.
It is also not a productivity tool. Treating the diary as a self-improvement scoreboard is how it stops working. Quantifying sleep can amplify orthosomnia (anxiety about sleep that worsens sleep), and the diary's value is observational, not prescriptive. Track for two weeks, look at the patterns, change one input at a time.
