Support vector machines for automated snoring detection: proof-of-concept

Area of Science:

• Sleep science
• Biomedical engineering
• Machine learning applications

Background:

• Snoring is linked to adverse health outcomes, independent of sleep disordered breathing.
• Objective snoring measurement is lacking in sleep studies due to quantification difficulties.
• Manual snoring scoring is time-consuming and labor-intensive, limiting study scope.

Purpose of the Study:

• To validate support vector machines (SVM) for automated snoring detection.
• To assess the feasibility of using machine learning for objective snoring quantification.
• To establish a scalable method for analyzing snoring in research.

Main Methods:

• Proof-of-concept study using a support vector machine (SVM) algorithm.
• Training and testing SVM on approximately 150,000 snoring/non-snoring data segments.
• Comparing SVM performance (F-score) against visual scoring using Wilcoxon signed rank test.

Main Results:

• SVM algorithm demonstrated comparable performance to trained visual scorers.
• No statistically significant difference in discriminating snore from non-snore segments (SVM F-score=82.46 ± 7.93 vs. visual F-score=88.35 ± 4.61, p=0.2786).
• SVM achieved reliable snoring classification.

Conclusions:

• The SVM algorithm is a viable tool for automated snoring detection.
• SVM performance is comparable to that of expert human scorers.
• Automated snoring detection using SVM can facilitate larger and multi-night sleep studies.