Snore sound recognition: On wavelets and classifiers from deep nets to kernels.

Area of Science:

• Biomedical Signal Processing
• Machine Learning in Healthcare
• Sleep Apnea Research

Background:

• Wavelet features are effective for snore sound classification.
• Previous studies did not clearly establish the benefits of Wavelet Transform Energy (WTE) and Wavelet Packet Transform Energy (WPTE).
• The impact of frame size and overlap on wavelet feature extraction requires further investigation.

Purpose of the Study:

• To comprehensively compare WTE and WPTE features for snore sound classification.
• To evaluate the influence of varying frame sizes and overlaps on wavelet feature extraction.
• To assess the performance of WTE and WPTE across multiple machine learning classifiers.

Main Methods:

• An updated snore sound database from 40 patients across three medical centers was utilized.
• Wavelet low-level descriptors were extracted, analyzing the effects of frame size and overlap.
• WTE and WPTE features were compared using Support Vector Machines (SVM), K-Nearest Neighbours (KNN), Linear Discriminant Analysis (LDA), Random Forests, Extreme Learning Machines (ELM), Kernel Extreme Learning Machines (KELM), Multilayer Perceptron (MLP), and Deep Neural Networks (DNN).

Main Results:

• When used with a Support Vector Machine (SVM) classifier, Wavelet Transform Energy (WTE) outperformed Wavelet Packet Transform Energy (WPTE) (p < 0.002).
• Wavelet Packet Transform Energy (WPTE) demonstrated significant improvement when trained with a K-Nearest Neighbours (KNN) classifier (p < 0.001).
• The study analyzed the impact of frame size and overlap on feature extraction, though specific results are detailed within the paper.

Conclusions:

• The choice of wavelet feature (WTE vs. WPTE) significantly impacts classification performance based on the chosen machine learning algorithm.
• WTE is optimal for SVM-based snore sound classification, whereas WPTE is more effective with KNN.
• This research provides valuable insights for optimizing feature selection in snore sound analysis for improved diagnostic accuracy.