Convolutional Neural Network-Based Obstructive Sleep Apnea Identification

Abstract

Obstructive Sleep Apnea (OSA) identification aims to recognize the sounds from the obstructive sleep apneahypopnea syndrome (OSANHS) patients. Despite remarkable advances have been made, the performance heavily relies on the sound representation. Feature selection is needed to improve the performance. Generally, the normal snoring and the snoring of OSANHS patients have a greater difference in acoustic characteristics. Ordinary snoring of human breathing is a regular, fluctuating and cyclical state, while OSANHS pathological snoring is often accompanied by a long pause. Based on the acoustic characteristics, this paper proposes an OSA recognition algorithm based on a convolutional neural network. First, the Mel-scale frequency cepstral coefficient (MFCC) of the sound are extracted. Then, convolutional neural network is deployed to predict the possibility of OSA. To empirically investigate the effectiveness and robustness of the proposed approach, extensive experiments were performed on a benchmark dataset. The obtained results showed that our method significantly outperforms related baselines and is also competitive or superior to the recently reported systems.