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

Objective: To validate automatic sleep stage classification using deep neural networks on sleep assessed by radar technology in the commercially available sleep assistant Somnofy® against polysomnography (PSG). Methods: Seventy-one nights of overnight sleep in healthy individuals were assessed by both PSG and Somnofy at two different institutions. The Somnofy unit was placed in two different locations per room (nightstand and wall). The sleep algorithm was validated against PSG using a 25-fold cross validation technique, and performance was compared to the inter-rater reliability between the PSG sleep scored by two independent sleep specialists. Results: Epoch-by-epoch analyses showed a sensitivity (accuracy to detect sleep) and specificity (accuracy to detect wake) for Somnofy of 0.97 and 0.72 respectively, compared to 0.99 and 0.85 for the PSG scorers. The sleep stage differentiation for Somnofy was 0.75 for N1/N2, 0.74 for N3 and 0.78 for R, whilst PSG scorers ranged between 0.83 and 0.96. The intraclass correlation coefficient revealed excellent and good reliability for total sleep time and sleep efficiency, while sleep onset and R latency had poor agreement. Somnofy underestimated total wake time by 5 min and N1/N2 by 3 min. N3 was overestimated by 4 min and R by 3 min. Results were independent of institution and sensor location. Conclusion: Somnofy showed a high accuracy staging sleep in healthy individuals and has potential to assess sleep quality and quantity in a sample of healthy, mostly young adults. More research is needed to examine performance in children, older individuals and those with sleep disorders