Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device
Level 3 - non-randomized controlled study
Diagnostic accuracy and algorithm validation cohort study against reference polysomnography; graded by design analogy.
OpenAlex W2937191485 · doi:10.1093/sleep/zsz180
What was done
Investigators collected raw acceleration and photoplethysmography heart rate data from Apple Watches worn by participants during polysomnography and an preceding ambulatory period using a custom application. They evaluated multiple machine learning classifiers incorporating motion, heart rate variability, and clock features for sleep-wake and sleep-stage prediction, and tested model generalizability on data from the Multi-Ethnic Study of Atherosclerosis (MESA).
What was found
Neural networks provided the best classification performance. For binary sleep-wake scoring, the model correctly classified 90% of epochs, achieving 93% sensitivity for sleep and 59.6% specificity for wake. Differentiating wake, NREM sleep, and REM sleep reached approximately 72% accuracy when using all features. Performance was comparable when tested on the external MESA dataset.
Why it matters
It demonstrates that raw sensor streams from consumer smartwatches can be analyzed with transparent, open algorithms to estimate sleep stages without relying on proprietary black-box software.
Limits
The abstract does not report participant sample size, demographics, or clinical characteristics. Specificity for wake detection was low (59.6%), and staging was restricted to three broad stages rather than standard five-stage polysomnography classification.
Cited by
- context Consumer sleep tracking devices have an accuracy of approximately 60% or lower when attempting to differentiate non-REM from REM sleep compared to polysomnography.