VOTE versus ACLTE: comparison of two snoring noise classifications using machine learning methods

Abstract

Background: Acoustic snoring sound analysis is a noninvasive method for diagnosis of the mechanical mechanisms causing snoring that can be performed during natural sleep. The objective of this work is development and evaluation of classification schemes for snoring sounds that can provide meaningful diagnostic support. Materials and methods: Based on two annotated snoring noise databases with different classifications (s-VOTE with four classes versus ACLTE with five classes), identically structured machine classification systems were trained. The feature extractor openSMILE was used in combination with a linear support vector machine for classification. Results: With an unweighted average recall (UAR) of 55.4% for the s‑VOTE model and 49.1% for the ACLTE, the results are at a similar level. In both models, the best differentiation is achieved for epiglottic snoring, while velar and oropharyngeal snoring are more often confused. Conclusion: Automated acoustic methods can help diagnose sleep-disordered breathing. A reason for the restricted recognition performance is the limited size of the training datasets.