Optimizing MFCC Parameters for Breathing Phase Detection

Area of Science:

• Biomedical Engineering
• Signal Processing
• Digital Health

Background:

• Breathing phase detection is crucial for clinical and digital health applications.
• Standard Mel Frequency Cepstral Coefficients (MFCCs) settings often yield suboptimal classification performance.
• There is a need for improved feature extraction methods for respiratory sound analysis.

Purpose of the Study:

• To systematically optimize Mel Frequency Cepstral Coefficients (MFCCs) parameters for enhanced respiratory phase detection.
• To evaluate the impact of MFCC parameter optimization on classification accuracy using Support Vector Machines (SVMs) and deep learning models.
• To identify a practical trade-off between accuracy and latency for real-time clinical applications.

Main Methods:

• A proprietary dataset of 1500 respiratory sound segments was utilized.
• Mel Frequency Cepstral Coefficients (MFCCs) parameters, including the number of coefficients, frame length, and hop length, were systematically optimized.
• Classification performance was assessed using Support Vector Machines (SVMs) and benchmarked against deep learning models (VGGish, YAMNet, MobileNetV2).

Main Results:

• Optimal MFCC parameters (30 coefficients, 800 ms frame length, 10 ms hop length) achieved 87.16% accuracy, a significant improvement over default settings (80.96%).
• Optimized MFCCs demonstrated performance equivalent to or better than established deep learning methods.
• A clinically practical frame length of 200-300 ms offered a favorable balance between accuracy (85.08%) and latency.

Conclusions:

• Optimized MFCC parameters substantially enhance the accuracy of respiratory phase classification.
• These optimized parameters provide efficient and interpretable solutions for real-time respiratory monitoring.
• Further validation in diverse clinical settings and exploration of advanced learning strategies are recommended.