A deep-learning approach to assess respiratory effort with a chest-worn accelerometer during sleep

The objective is to develop a new deep learning method for the estimation of respiratory effort from chest-worn accelerometer during sleep. We evaluate performance, compare it against state-of-the art method, and assess whether can differentiate between sleep stages. In 146 participants undergoing overnight polysomnography data were collected an worn on chest. study partitioned into train, validation, holdout (test) sets. used train validation sets generate convolutional neural network performed model selection respectively, while we set (72 participants) performance. A with 9 layers 207,855 parameters was automatically generated trained. significantly outperformed best performing conventional based Principal Component Analysis; reduced Mean Squared Error 0.26 0.11 also better in detection breaths (Sensitivity 98.4 %, PPV 98.2 %). addition, exposed significant differences characteristics stages (p < 0.001). predicts low error sensitive precise breaths. reproduces stages, which may enable automatic staging, using just accelerometer.