Deep Neural Network Sleep Scoring Using Combined Motion and Heart Rate Variability Data

Area of Science:

• Sleep Science
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Wrist actigraphy's sleep assessment accuracy relies on both hardware and interpretative algorithms (IAs).
• Existing IAs have limitations in precisely evaluating sleep quality.
• Deep learning offers potential for developing more sophisticated sleep analysis tools.

Purpose of the Study:

• To develop a novel deep learning-based interpretative algorithm (IA) for wrist actigraphy.
• To improve sleep quality assessment compared to existing IAs.
• To integrate activity count and heart rate variability (HRV) metrics for enhanced sleep scoring.

Main Methods:

• Utilized simultaneous polysomnography (PSG) and wrist actigraphy data from 222 participants.
• Applied deep learning models using activity count and HRV metrics with varying window lengths (30s, 3min, 5min).
• Developed a novel Haghayegh Algorithm (HA) and compared its performance against UCSD and Actiwatch proprietary IAs.

Main Results:

• The HA model, using activity count and 5-min window HRV, showed the highest agreement with PSG.
• HA achieved 84.5% accuracy, outperforming comparator IAs by 5.3-6.2%.
• HA demonstrated superior specificity and Kappa agreement for sleep epoch detection and comparable performance in deriving key sleep parameters.

Conclusions:

• The novel HA model significantly enhances sleep scoring performance over existing popular IAs.
• Simultaneous use of activity count and 5-min window HRV metrics is key to HA's improved accuracy.
• This deep learning approach represents a significant advancement in objective sleep assessment using wrist actigraphy.