Validation of sleep stage classification using non-contact radar technology and machine learning (Somnofy®)

Area of Science:

• Sleep Science
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Polysomnography (PSG) is the gold standard for sleep assessment.
• Automatic sleep stage classification using deep neural networks offers a potential non-invasive alternative.
• Radar technology, as used in devices like Somnofy, presents a novel approach to sleep monitoring.

Purpose of the Study:

• To validate the automatic sleep stage classification accuracy of the Somnofy device against PSG.
• To assess the performance of deep neural networks in radar-based sleep analysis.
• To compare Somnofy's performance to inter-rater reliability of human PSG scoring.

Main Methods:

• Seventy-one nights of sleep data from healthy individuals were collected using both PSG and Somnofy.
• Somnofy's sensor placement was varied (nightstand and wall) across two institutions.
• A 25-fold cross-validation technique was employed to validate the Somnofy algorithm against PSG.

Main Results:

• Somnofy demonstrated high sensitivity (0.97) and specificity (0.72) for sleep/wake detection compared to PSG scorers.
• Accuracy for specific sleep stages (N1/N2, N3, REM) ranged from 0.74 to 0.78, with good reliability for total sleep time and sleep efficiency.
• The device showed minor under/overestimations for certain sleep stages and wake time, with results independent of institution and sensor location.

Conclusions:

• Somnofy exhibits high accuracy in staging sleep for healthy individuals, indicating potential for assessing sleep quality and quantity.
• The findings suggest Somnofy is a promising tool for sleep monitoring in healthy, predominantly young adults.
• Further investigation is recommended to evaluate Somnofy's performance in children, the elderly, and individuals with sleep disorders.