A neural network technique for feature selection and identification of obstructive sleep apnea

A novel identification method of Obstructive Sleep Apnea from normal controls is presented in this paper. The method uses the approximate power spectral density of heart rate variability, which is estimated using a soft-decision wavelet-based decomposition in a combination with a neural network. The neural network is used for two purposes: to select the optimum frequency bands that can be used for identification during the feature extraction step, and to identify the data during the feature matching step. Two sets of data, training set and test set, which are downloaded from the MIT-data bases, are used in this work. The training set, which consists of 20 obstructive sleep apnea subjects and 10 normal subjects, is used to train the neural network of type feed-forward back-propagation. The test set, which consists also of 20 obstructive sleep apnea and 10 normal subjects is used to test the performance of the identification system. A best identification efficiency of 93.33% has been obtained in this work using three inputs only.