Unsupervised classification of respiratory sound signal into snore/no-snore classes

Characterisation of Upper Airway Collapse in OSA Patients Using Snore Signals: A Cluster Analysis Approach

This paper provides the results of an unsupervised learning algorithm that characterize upper airway collapse in obstructive sleep apnoea (OSA) patients using snore signal during hypopnoea events. Knowledge regarding the site-of-collapse could improve the ability in choosing the most appropriate treatment for OSA and thereby improving the treatment outcome. In this study, we implemented an unsupervised k-means clustering algorithm to label the snore data during hypopnoea events. Audio data during sleep were recorded simultaneously with full-night polysomnography with a ceiling microphone. Various time and frequency features of audio signal during hypopnoea were extracted. A systematic evaluation method was implemented to find the optimal feature set and the optimal number of clusters using silhouette coefficients. Using these optimal feature sets, we clustered the snore data into two. Performance of the proposed model showed that the data fit well in two clusters with a mean silhouette coefficients of 0.79. Also, the clusters achieved an overall accuracy of 62% for predicting tongue/non-tongue related collapse.

SnoreNet: Detecting Snore Events from Raw Sound Recordings

Snoring is one of the earliest symptoms of Obstructive Sleep Apnea Hypopnea Syndrome (OSAHS). Snore detection is the first step in developing non-invasive, low-cost, and totally sound-based OSAHS analysis approaches. In this work, we propose a simple yet effective deep neural network, named SnoreNet, for detecting snores from a continuous sound recording. Without manually crafted features, SnoreNet can capture the characteristics of snores. Since snore varies in temporal length, SnoreNet combines output from multiple feature maps to detect snore. In each feature map, SnoreNet uses a set of default bounding box generated by a base length and different scales to match snores. SnoreNet adjusts the box to better locate snores and predicts a score for the presence of snore in each default bounding box. The performance of SnoreNet was evaluated on a newly collected snore pattern classes dataset, which achieves 81.82% average precision (AP).

Research on Snoring Recognition Algorithms

A snoring recognition algorithm based on machine learning is proposed to effectively and precisely recognize snoring. To obtain a dataset, the speech endpoint detection algorithm and Mel frequency cepstrum coefficient feature extraction algorithm are applied to process speech signal samples. The dataset is classified into snoring and nonsnoring data (other speech signals) using support vector machines. Experimental results show that the algorithm recognizes snoring signals with a high accuracy rate of 97% and positively impacts subsequent research and related engineering applications.