Sleep Tech Lab.

Lab Methodology

Marketing copy doesn't survive contact with clinical data. This is the exact protocol we use to strip away the noise and derive the actual Pearson correlation coefficients for consumer sleep tech.

1. The Baseline: Why EEG matters

Sleep is fundamentally a neurological process, not a cardiovascular one. The brain moves through distinct electrical frequencies—from the rapid, low-voltage beta waves of wakefulness, down to the slow, high-amplitude delta waves (0.5 to 2 Hz) that define deep N3 sleep.

Because consumer devices (like the Apple Watch or Oura Ring) cannot read brainwaves, they rely on Photoplethysmography (PPG) and accelerometers. They are guessing your brain state based on your heart rate variability (HRV) and movement. To test if they are guessing correctly, we must compare them to an actual brainwave sensor.

We utilize the Dreem 2 Headband as our ground-truth baseline. It is a consumer-available, 5-dry-electrode EEG device that has been independently validated against clinical polysomnography (PSG) in multiple peer-reviewed studies with a >83% epoch-by-epoch agreement.

2. The Data Collection Protocol

  • Duration: A minimum of 14 consecutive nights per device tested. This ensures sufficient data volume (approx. 100+ hours of sleep architecture) to account for natural variations in sleep latency and cycle length.
  • Placement: Wrist-worn devices are placed on the non-dominant wrist to minimize false movement artifacts. Rings are placed on the index or middle finger.
  • Concurrent Recording: The test device and the Dreem 2 EEG are worn simultaneously. Both devices are fully charged and synced to their respective servers every morning.

3. Epoch Analysis & Alignment

Clinical sleep scoring divides the night into 30-second segments called epochs. To compare a wearable to our EEG baseline, we must align these epochs perfectly.

Data Handling: We export the raw CSV files from both the test device API and the baseline device. We utilize a custom Python script to perform time-series alignment, snapping timestamps to the nearest 30-second boundary. If there is a data gap (e.g., the sensor lost contact with the skin), that specific epoch is excluded from the final calculation rather than interpolated, to prevent artificial inflation of accuracy.

4. Metrics Computed

We do not use proprietary "Sleep Scores". We calculate the following mathematical agreements:

  • Total Sleep Time (TST) Error: The absolute difference in minutes between the baseline and the test device.
  • Wake After Sleep Onset (WASO) Sensitivity: The ability of the device to detect brief awakenings during the night. (Most wearables fail spectacularly here, mistaking lying still for sleeping).
  • Pearson Correlation Coefficient (r) for Stages: We isolate Light, Deep (N3), and REM sleep, and calculate the correlation coefficient against the baseline. An r value of 1.0 means perfect agreement; 0.0 means no relationship. Consumer devices rarely exceed r=0.8 for specific sleep stages.

Frequently Asked Questions

Why 14 nights?
A single night of data is statistically useless due to physiological anomalies (e.g., alcohol consumption, stress, temperature). 14 nights provides a large enough sample size to smooth out daily variance and reveal the algorithm's true baseline accuracy.
Do you accept sponsored devices?
No. We purchase all test units at retail price to ensure we are testing the exact hardware and firmware available to consumers, free of manufacturer selection bias.

Further Reading