The short answer
For estimating sleep stages, using only Apple Watch heart rate performed about as well as using heart rate variability taken from clinical PSG (agreement κ=0.221 against 0.244). But both sit at "moderate agreement", not accuracy. Meanwhile the nightly summary figures — the share of REM sleep, sleep efficiency — barely correlated with the PSG values at all. The shape of the night is worth reading; the "REM 22%" an app hands you is not a number to take at face value.
Key points
- Apple Watch heart rate alone matched clinical PSG-derived HRV at estimating stages
- But agreement sits around κ=0.22, which is not the same as being accurate
- Remove the surrounding time window and κ falls to 0.09 — a single moment of heart rate is not enough
- Correlation with REM% was r=0.003. Reproducing nightly summaries missed the target entirely
- Swapping cohorts barely changed performance (AUROC fell 0.028), so this is not overfitting
What was compared with what
The ground truth for sleep stages is polysomnography (PSG), which records brain waves, eye movement and muscle activity together. It needs a technician for the night and more than twenty electrodes on the body. That is the reference.
Against it, we compared two approaches under identical conditions. One used heart rate variability derived from clinical PSG (24 features). The other used only the heart rate available from an Apple Watch (11 features), evaluated on data published by Walch and colleagues in which an Apple Watch and PSG recorded the same nights (25,813 epochs).
Result 1: the wrist held its own against clinical data
The gap is 0.023 in κ. Against the intuition that a wrist cannot possibly be enough, using clinical-grade heart rate variability barely changed the result — at least for estimating stages.
That is a point in the watch's favour, and equally a limit on the whole approach. κ=0.22 counts statistically as moderate agreement, which is what reading sleep stages from the heart without an EEG looks like.
- Clinical PSG heart rate variability (24 features): macro AUROC 0.711, agreement κ 0.244
- Apple Watch heart rate only (11 features): macro AUROC 0.708, agreement κ 0.221
Result 2: strip out the surrounding minutes and it collapses
Recomputing the same method with the surrounding ±5 epochs (about five and a half minutes) removed dropped performance sharply.
- Clinical PSG heart rate variability: κ 0.244 → 0.119
- Apple Watch heart rate only: κ 0.221 → 0.092
Result 3: nightly summaries could not be reproduced
This is the important part. If stages can be estimated, it seems reasonable that summary figures — sleep efficiency, REM% — would follow. We set the bar in advance at a correlation of r≥0.85 against PSG. Across 67 recordings:
- Wake after sleep onset (WASO): r = 0.42
- Share of deep sleep (N3%): r = 0.32
- Sleep efficiency: r = 0.30
- Sleep latency: r = −0.007
- Share of REM sleep: r = 0.003
So what is worth looking at
What follows is that reading wearable sleep data as a nightly report card does not hold. If REM% is five points lower than yesterday, there is no way to tell whether your sleep changed or the estimate wobbled.
Feelmo does not put sleep stages front and centre, and treats them as a trend across nights, precisely because of this. Your own record, gathered under similar conditions and read across days and weeks — for now that is the honest use of a sensor on a wrist.
Conditions and limits
- Holding out a whole cohort and retraining cost only 0.028 in AUROC, so this is not fitted to one dataset
- Everything was evaluated on public datasets; no Feelmo user data was involved
- The nightly-summary test used 67 recordings, so the correlation values carry a wide interval
- This tests the accuracy of stage estimation. It does not determine whether a sleep disorder is present