Dual Pressure Signal Internal Feature Based Obstructive Sleep Apnea Detection

Obstructive sleep apnea (OSA) is one of major sleep disorders, which has been considered as a high-risk factor of cardiovascular disease, stroke, and diabetes. Polysomnography (PSG) is the gold standard for OSA detection. But it has to be conducted in a specified clinical setting requiring trained physicians, with high cost, time consuming, and laborious. In this paper, we develop an automated OSA detection method based on pressure signals. Specifically, we use the piezoelectric ceramic sensors array to collect pressure signals generated by subjects during sleep. To capture the temporal-frequency feature and internal temporal conversion feature from the collected pressure signals, we use transformation methods to obtain time-frequency images and Markov Transition Field (MTF) images, respectively. To fully learn the feature representation in a different view, we propose a fusion neural network with ResNet18 and ResNet50 for OSA detection. Experiment results on 36 overnight signals collected from Peking Union Medical College Hospital showed that our proposed method can achieve 80.00% sensitivity, 69.63% specificity, 74.81% accuracy, 72.48% precision, and 0.76 F
1
score, which is competitive with existing state-of-the-art methods. It means that the proposed method can potentially provide sleep monitoring service in a home setting and aid in future medical diagnoses.