SpO2 based Sleep Apnea Detection using Deep Learning

In a classical classification process, automatic sleep apnea detection involves creating and selecting the features, using prior knowledge, and apply them to a classifier. A different approach is applied in this paper, where a Deep Belief Network is used for feature extraction, without using domain-specific knowledge, and then the same network is used for classification of sleep apnea. The Deep Belief Network was created by stacking Restricted Boltzmann Machines. The first two layers are autoencoder type and the last layer is of soft-max type. The initial weights are calculated using unsupervised learning and, at the end, a supervised fine-tuning of the weights is performed. Two public databases, one with 8 subjects and other with 25 subjects, are tested using tenfold cross validation. The optimum number of hidden neurons of this problem is found using a search technique. The accuracy achieved from UCD database is 85.26% and Apnea-ECG database is 97.64%.