Research On The Source Classification Method Of Snoring Based On Microphone Array

Snoring is a prevalent sleep-related disorder including simple snoring and obstructive sleep apnea/hypopnea syndrome(OSAHS).As the primary indicator of snoring,loud snoring sounds will decrease sleep quality of patients and their roommates,leading to interpersonal conflicts.In addition,OSAHS increases the risk of cardiocerebral vascular diseases.Therefore,the treatment of snoring has important significance and the information of snoring source is essential for a targeted treatment decision.Drug-induced sleep endoscopy(DISE)is a widely used method to determine the snoring source at present.However,DISE is expensive,timeconsuming and invasive.Considering that snoring sounds carry significant information about tissue vibrations within the upper airway,analyzing snoring sounds becomes a promising means for identifying the snoring source,which is easy to generalize.Based on the National Natural Science Foundation of China(61271410),in this work we propose a method of snoring source classification.Environmental noise,which is inevitably received by microphones when collecting snoring sounds,will leads to the reduction of classification performance.Therefore,we apply a generalized sidelobe canceller-based adaptive enhancement technique to depress the noise.Then we extract five categories of features including specific frequencies,energy ratios,fundamental frequency and formants,MFCC,and wavelet energy features.Besides,we propose a new feature named compressed histogram of oriented gradients(CHOG)to characterize the spectral energy information,which is obtained by extracting histogram of oriented gradients(HOG)from spectrograms and then compressed via the multilinear principal component analysis algorithm.In order to obtain a more discriminative feature set,we apply the ReliefF algorithm to the aggregation of all features.Finally,a support vector machine(SVM)is used to classify the snoring sources.The experiments are conducted using the snoring events from 76 patients and the results show that CHOG is superior to other single features with an average classification accuracy of 89.8% and other higher metrics.In addition,the fusion feature set achieves higher classification accuracy and better stability,which confirms that multi-feature analysis and feature selection can improve classification performance.All the results above show that the proposed method is effective for snoring source classification with important medical and social value.