Automatic classification of snoring sounds from excitation locations based on prototypical network

Abstract

Snoring site is greatly associated with the treatment approach of Obstructive Sleep Apnea (OSA), a serious chronic sleep disorder. There are more demands to recognize snoring sounds automatically from different excitation locations with high accuracy in clinical practice. Based on the original division of Munich Passau Snore Sound Corpus (MPSSC) with limited four types of snoring sounds. A meta-learning algorithm named prototypical network embedded with a convolutional neural network (CNN) in the prototypical network is used to classify these snoring sounds. The result shows that CNN with six-layer convolution and complement-cross-entropy loss function could well solve the problem of imbalance distribution of the MPSSC dataset to yield the highest unweighted average recall with the value of 78.85% in the development set and 77.13% in the test set under all test conditions. Compared with the challenge baseline, this paper yields a high improvement with a value of 18.63% (p < 0.05) in the test set. It indicates that the simplicity and effectiveness of the prototypical network make it a promising approach for biological signal classification with a relatively small dataset.