Comparison of PSG signals and Respiratory Movement Signal via 3D Camera in Detecting Sleep Respiratory Events by LSTM Models

A contactless system to detect respiratory events during sleep may be advantageous because it enables normal sleeping pattern and eliminates the need for constant monitoring of contact sensors. To detect sleep respiratory events, a 3D timeof-flight (TOF) camera is placed above the bed to measure a respiratory movement signal. Using this signal, we trained a long-short term memory (LSTM) model to detect respiratory events. In addition, we trained LSTM models based on SpO2, and abdomen and thorax respiratory inductance plethysmography (RIPsum). LSTM models were trained on 8 folds using 61 synchronized 3D video and polysomnography (PSG) recordings of patients with suspected sleep apnea to classify 30-second segments as either respiratory event or normal breathing. Manual PSG annotations served as ground truth. The LSTM model based on 3D TOF camera achieved a mean accuracy of 0.79 and mean Cohen’s kappa of 0.54. SpO2 based model scored a mean accuracy of 0.86 and mean Cohen’s kappa of 0.68 while RIPsum based model scored a mean accuracy of 0.82 and mean Cohen’s kappa of 0.61. The 3DRespMvt performance can be improved by combining it with SpO2, resulting in a mean accuracy of 0.87 and mean Cohen’s kappa of 0.71.