Snoring classification with deep time-frequency features.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Snoring is a key indicator of Obstructive Sleep Apnea (OSA).
• Accurate snore classification is vital for non-invasive localization of upper airway obstructions.
• Current methods face challenges with limited data, class imbalance, and poor time-frequency integration.

Purpose of the Study:

• To develop a heterogeneous framework for improved snore signal classification and upper airway obstruction localization.
• To address data limitations and enhance the integration of time-frequency information in snore analysis.

Main Methods:

• A novel framework combining Short-Time Fourier Transform (STFT), pre-trained Convolutional Neural Networks (CNNs), and L2-regularized Support Vector Machine (SVM).
• STFT converts snore signals into time-frequency spectrograms, preserving crucial signal structures.
• Deep features extracted using a pre-trained AlexNet mitigate limited labeled data issues.
• An L2-regularized SVM classifier is employed to prevent overfitting in high-dimensional, small-sample scenarios.

Main Results:

• The proposed method achieved a test set Unweighted Average Recall of 67.1% on the Munich-Passau Snore Sound Corpus.
• Outperformed existing state-of-the-art methods, including end-to-end CNNs and Transformer-based audio models.
• Ablation studies showed significant performance drops (up to 21.3%) when any component (STFT, pre-trained CNN, SVM) was removed, confirming their importance.

Conclusions:

• The developed heterogeneous integration framework offers an effective, generalizable, and data-efficient solution for snore source localization.
• This approach significantly improves the classification of snoring sounds for diagnosing Obstructive Sleep Apnea.
• The findings highlight the benefit of combining STFT, pre-trained CNNs, and SVM for complex audio signal analysis in medical applications.