La vérité sur le suivi du sommeil des montres connectées

Main ideas / concepts conveyed

• Sleep trackers (smartwatches/rings/bracelets) can look convincing but may be statistically misleading.

  • Example claim: a “90/100 sleep score” and colorful graphs (deep/REM/light phases) can be wrong in a significant portion of cases due to estimation limits.

• Two different tasks are involved in wearables:

  1. Detecting when you fall asleep and when you wake up (sleep duration).
  2. Estimating sleep stages/phases (deep sleep, REM, light sleep, and wakefulness).

• Sleep duration detection is relatively decent; sleep phase detection is much less reliable.

  • Wearables determine sleep onset using proxies such as stillness, body temperature drop, and heart-rate decrease.
  • They work reasonably well for total sleep time, with only minor issues (e.g., if you remain motionless in bed while reading/watching).

• Sleep scores (0–100) are computed from estimated stages.

  • The score is described as based on:
    • light sleep
    • deep sleep
    • REM sleep
    • wakefulness phases
  • Multiple brands/devices are referenced as offering such scoring (e.g., Samsung, Garmin, Hu(a)??, Apple).

• Devices rely on indirect signals rather than brain-wave measurement.

  • Gold standard: polysomnography (PSG), which uses an electroencephalogram (EEG) to directly infer sleep stages.
  • Wearables do not use EEG, so they infer phases using:
    • motion/micro-movements
    • temperature changes
    • heart-rate patterns

• Sleep-stage classification by wearables has notable accuracy limitations.

  • A cited meta-analysis (last year) analyzing 24 studies, 800 participants:
    • Wearables underestimate total sleep time by ~17 minutes on average (~4.7%).
    • Framed as “not too bad” for duration.
  • The “real problem” is distinguishing sleep phases.
  • A cited 2024 large study (based on 35 articles) estimates:
    • sleep phase accuracy around 69–79%
    • meaning in 20–30% of cases, sleep phases are poorly analyzed (i.e., often wrong)
    • roughly: ~7 times out of 10 correct, but ~3 times out of 10 wrong

• Even the “most reliable” consumer devices are not reliably accurate for sleep staging.

  • Examples listed as relatively more reliable:
    • connected rings (from Oura)
    • bracelets (from Whoop)
    • Fitbit watches/bracelets (Fitbit is described as belonging to Google)
  • Even in best-case scenarios (example: Oura referenced as “Hura tanks”), reliability is claimed to be around ~80%.

• Why it’s hard (and risky) to verify wearable accuracy in practice

  • To validate sleep-stage accuracy, you’d need polysomnography (claimed cost: ~€20,000).
  • The speaker also claims a commonly used consumer reference EEG head device (Dream 2) is only ~85% reliable, making it difficult to objectively evaluate wearables against a “true” gold standard.

• What wearables can still be useful for: trends, not medical diagnosis

  • The argument: sleep trackers are often inaccurate on absolute values (over/underestimating).
  • But the trend over time can be meaningful—compared to:
    • a smart scale (useful for weight trends)
    • a step counter (may miscount incidental movement, but helps for long-term patterns)
  • Therefore: use sleep scores for personal trend awareness, not for diagnosing disease.

• Practical takeaway

  • If you wake up still tired, the speaker suggests you likely didn’t sleep enough and you should not trust the displayed high score (e.g., “95/100”) as a guarantee.

• Video closing

  • Mentions an article in the description and that sources for the cited studies are provided.

---

Methodology / instructions presented

How sleep watches generally estimate sleep onset/offset (as described)

• Detect that you are completely still
• Observe slight drop in body temperature
• Observe heart-rate decrease
• Conclude you are falling asleep
• Determine sleep duration from the inferred sleep interval
• Note: if you stay motionless awake in bed (reading/watching), it can be mistaken as sleep

How sleep phase estimation is inferred (as described)

• Use indirect signals instead of EEG:
  • micro-movements
  • heart-rate levels and variability
  • temperature
• Example heuristic:
  • very still + very low heart rate → interpreted as deep sleep
  • very still + slightly higher and more variable heart rate → interpreted as paradoxical/REM-like sleep
• Critical caveat:
  • these heuristics can fail (e.g., deep sleep can still occur while still and with a low heart rate; feelings of restfulness may confound interpretation)

How to interpret sleep scores in a useful way (recommended approach)

• Don’t use sleep scores as absolute truth or medical diagnostics
• Focus on directional change over days/weeks
• Use long-term patterns (trends) rather than single-night numbers

Decision rule the speaker suggests

• If you wake up tired, treat that as a stronger signal than the smartwatch’s high score.

---

Speakers / sources featured

Speakers

• The video narrator/speaker (not named in the provided subtitles)

Sources (studies and technical references mentioned)

• Meta-analysis (published last year)
  • 24 studies, 800 participants
• Large study published in 2024
  • 35 articles
• Polysomnography (PSG) / electroencephalogram (EEG) as the gold standard
• Dream 2 (EEG-based head device referenced as an imperfect reference)
• Device brands referenced (examples of products that provide sleep scoring/tracking):
  • Samsung
  • Garmin
  • Apple (Watch/WatchOS, iOS)
  • Withings (sleep sensors)
  • Oura (connected ring)
  • Whoop (connected bracelet)
  • Fitbit (bracelets/watches; described as belonging to Google)
  • Hura (appears to be referenced alongside Oura; exact name may be an auto-subtitle error)
  • Indexli Monitor (described as an armband from Garmin; subtitle may contain errors)