Automatic detection of snoring signals: validation with simple snorers and OSAS patients

Relationship between snoring and Obstructive Sleep Apnea Syndrome (OSAS) has been reported in the literature. Recently, studies of snoring sound intensity, but also estimation of spectral features for each snoring episode, have been published. Usually, patients that are suspected of OSAS pathology are studied by polysomnography during all the night. To analyze the snoring signal, it is very useful to automatically detect each episode, in order to calculate several features that describe the signal. In this work an automatic detection algorithm of acoustic snoring signals has been designed, to work with long duration respiratory sound recordings. Two blocs compose the detector. The former is a segmentation subsystem that detects changes of variance on the signal. The latter is a 2-layer Feedforward Multilayer Neural Network with backpropagation learning algorithm. The network was trained with 625-selected events, including snores with different shapes and characteristics, from normal snorers and OSAS patients, and other sounds. In this way, the detector was designed to select snoring episodes from simple snorers and OSAS patients, and to reject cough, voice and other artifacts. The detector has been applied to real snoring signals recorded during polysomnographic studies. In order to validate the detector, more than 500 snores were analyzed from 10 excerpts, taken at random from a database of 30 snorer subjects with different apnea/hipoanea index (AHI). Results were compared with manual annotations done by a medical doctor. The detector showed a good performance and achieved a Sensitivity of 82% and a Positive Predictive Value of 90%.