Smartwatch based respiratory rate estimation during sleep using CNN/LSTM neural network

Breathing pattern abnormality during sleep leads to disturbance of sleep stages pattern and even to serious sleep disorders, like apnea. The breathing pattern is typically described by the respiratory rate variability. State of the art methods of respiratory rate measuring use thermal sensor placed in nasal channels or elastic chest belt. Both these methods are not appropriate for comfortable unobtrusive sleep. Here, we present a neural network method for respiratory rate estimation using photoplethysmogram signal. We propose to combine convolutional and recurrent approach. The algorithm was developed and validated on a combination of public clinical database of 8 subjects and private database of 96 subjects. The proposed model provides an average respiratory rate estimation error lower than 2.2 breaths per minute, and is applicable for implementation in wearable smart devices.