Obstructive Sleep Apnea Detection Using Combination of CNN and LSTM Techniques

This paper proposed a novel DL method for OSA detection. RR intervals (RRIs) was extracted from Apnea ECG signals at the first stage, which reduced computational complexity and highlighted OSA behavior. Then a model combining CNNs with LSTMs was proposed, which contained two continuous convolutional layers and an unfolded bidirectional LSTM. Apnea-ECG datasets from physionet.org were used for model train and test. Nesterov accelerated gradient (NAG) algorithm was applied to accelerated the convergence speed of network. The model achieved a specificity and sensitivity and of 96.94% and 98.97%, respectively and the accuracy was up to 99.80% at the same time, which produced better performance in detecting OSA compared with traditional neural network methods.