Deep learning algorithm for automated sleep stage scoring using heart rate variability

Sleep evaluations currently require multimodal data collection and manual review by human experts. In this study, we applied deep learning methods to create an algorithm for automated sleep stage scoring using ECG derived heart rate. The resultant algorithm was validated on 2 large, independent polysomnography datasets with 800 and 993 subjects. The model has an overall performance of 0.76 accuracy and 0.65 kappa against the reference stages labeled by AASM licensed experts classifying every 30 sec of sleep into 4 classes - WAKE, Light sleep, Deep sleep and REM. To our knowledge, this is the best reported performance of automated staging without EEG, especially on a dataset of this magnitude. Moreover, we demonstrate that the algorithm performance generalizes very well to new datasets which were never exposed to the training or validation process, including on subjects with and without apnea. We hope our results encourage further research necessary to determine the feasibility of applying this algorithm in the clinical setting and whether use of the algorithm could lead to improved care and outcomes and assist in identifying undiagnosed sleep disorders.

• Machine Intelligence

• Health & Bioscience