Sleep Apnea Classification Using Deep Neural Network

Abstract

At present, sleep related disorders are very common among the population, due to the varying stress and living habits of the individuals. Obstructive sleep apnea (OSA) is one such serious sleep disorder, where the person experiences a breath cessation while sleeping. An improper or any delayed diagnosis of this disorders can cause severe health issues. This work instigates a sleep apnea classification system using deep neural networks (DNN) using the Heart Rate Variability (HRV) and Respiratory Variability (RRV) features. As an initial step towards the work, a baseline system was developed using the statistical features using SVM as the backend classifier. The dataset was segmented into samples of 2 minutes, the features were extracted from the database, and were given to SVM model which showed an overall accuracy of 62.70% absolute. The patient and stage specific features seen in the PSG data are removed using a feature normalization technique called Covariance Normalization (CVN). Further a deep neural network system with 4 hidden layers were developed using the CVN performed features and it performed with an overall accuracy of 88.03% absolute on the training set and 84.21% absolute on the test set..