Adversarial Transfer Learning Architecture for Sleep Apnea Detection and Classification

Sleep Apnea is a common sleep disorder which causes repeated interruption in their breathing during sleep. Sleep apnea detection and classification is a vital task for identification of Sleep quality that are prone to hypopnea and Sleep apnea events which further lead to changes in the immune system especially on the various memory and metabolism properties. Sleep Apnea can be diagnosed using invasive or non-invasive approaches. In Particular, single channel ECG signals has been used in the large studies as it is processed by dividing the signal into 30 second segments containing complex time characteristics and frequency characteristics. However, it makes it difficult for manual techniques to discriminate between the different types of sleep apnea. In order to address those challenges, automated sleep apnea detection and classification model using deep learning mechanism has to be proposed. In this article, two adversarial transfer learning has been framed as prediction function to detect and classifies the type of Sleep apnea with AHI (Apnea-Hypopnea Index). Initially ECG derived respiration signal considered as EDR is extracted from PSG dataset through signal processing methods. In this work, Gabor filtering is used for noise reduction and Pan-Tompkins technique is used to determine R peaks in the QRS region of the noise eliminate signal. Next, minute by minute processing of RR interval is used to the extract crucial features for sleep apnea detection. Extracted feature is processed using the particle swarm optimization to select the optimal features. Finally adversarial neural network composed of multiple layers is used for detection of the sleep apnea on processing the source apnea features and target sleep apnea features in form map. Feature map with AHI is used to classify the sleep apnea into low, mild and severe API in SoftMax layer. We used approximately 100 patches of signals from different patients extracted from a SC4001E0-PSG dataset and 80% of dataset employed to train a model and 20% of dataset is used to validate the proposed model on cross fold validation. The results of K fold cross-validation experiments show that the proposed model achieves correlation of about 0.90 with the estimation of multiple sleep stages as highly automated model for characterization of sleep apnea.