AI-Driven sleep staging from actigraphy and heart rate

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Medicine

Background:

• Polysomnography (PSG) is the gold standard for sleep studies but is costly and inconvenient.
• Wrist-worn wearables offer a promising, accessible alternative for sleep monitoring.
• Current wearable sleep staging is limited to two classes, lacking detailed sleep insights.

Purpose of the Study:

• To develop and validate an AI technique for multi-class sleep staging using consumer-grade wearables.
• To enable accurate estimation of sleep stage durations, particularly deep sleep.
• To overcome data limitations of wearables compared to clinical PSG.

Main Methods:

• Developed Sequence-to-Sequence LSTM for Automated Mobile Sleep Staging (SLAMSS).
• Utilized wrist-accelerometry and heart rate data for three-class (wake, NREM, REM) and four-class (wake, light, deep, REM) staging.
• Validated SLAMSS on two large cohorts (MESA, MrOS) and an independent Apple Watch dataset.

Main Results:

• SLAMSS achieved high accuracy in three-class (up to 79%) and four-class (up to 72%) sleep staging.
• The model accurately predicted sleep stage durations, including underrepresented deep sleep.
• Validation on independent datasets and an Apple Watch dataset confirmed model robustness.

Conclusions:

• SLAMSS provides accurate multi-class sleep staging from feature-poor wearable data.
• Accurate deep sleep estimation from wearables holds significant potential for remote health monitoring.
• This AI-driven approach can facilitate long-term sleep health assessment and disease prediction.