Classification of pediatric snoring episodes using deep convolutional neural networks

Simple snoring and obstructive sleep apnea are health conditions that have psychological, physiological and social implications for children including depression, reduced cognitive ability, and hyperactive behavior. According to studies, the frequency of diagnosis of OSA and SS in children are given as 4.9%-17.1%. Obstructive sleep apnea is characterized by drops in blood oxygen level caused by blockage of the upper respiratory tract. A 50% drop in respiration for 10 seconds or more and a consequent 3% drop in blood oxygen level is defined as hypoapnea. Complete stopping of respiration for 10 seconds or more is defined as apnea. Apnea-Hypoapnea Index (AHI) is a commonly used unit for determining the severity of the apnea. For a diagnosis of OSA, an AHI greater than 5 for adults and an AHI greater than 1 for children are used as thresholds. Diagnosis of obstructive sleep apnea requires expensive and time-consuming measurements to be made in a sleep laboratory, hence the current approach is not feasible for screening at large scale. A low-cost system that is able to warn about signs of obstructive sleep apnea can provide a feasible alternative. In this work, a deep convolutional neural network is designed for automatic detection of snoring episodes of children from ambient audio data. In order to test the performance of the network, a dataset collected from 38 pediatric patients with simple snoring or obstructive sleep apnea. The ambient audio signal is converted to Mel spectrogram for the model to learn relationships between features in the frequency domain. The network is trained using Adadelta algorithm on 1.200 samples. In order to fine-tune the system parameters, a validation set of 300 samples is used. Finally, a test set of 300 samples was used observe the performance of the developed network. The samples in the training, validation, and test samples are all exclusive. The mean accuracy of the proposed system over 20 experiments is observed as 91% for snore and non-snore. The accuracy observed is higher than previous work on the same dataset.