A temporal dependency feature in lower dimension for lung sound signal classification

Area of Science:

• Medical Signal Processing
• Bioacoustics
• Machine Learning for Healthcare

Background:

• Respiratory sounds are complex, nonlinear, and nonstationary signals.
• Traditional Mel-frequency cepstral coefficients (MFCCs) lose temporal information and operate in high-dimensional spaces.
• Accurate classification of lung sounds is crucial for diagnosing respiratory conditions.

Purpose of the Study:

• To propose a novel feature extraction method, shifted delta-cepstral coefficients in lower-subspace (SDC-L), for lung sound classification.
• To evaluate the effectiveness of SDC-L in preserving temporal dependencies and reducing dimensionality.
• To compare SDC-L performance against existing methods using various machine learning and deep learning models.

Main Methods:

• Developed shifted delta-cepstral coefficients in lower-subspace (SDC-L) by modifying the Empirical Mode Decomposition (EMD) algorithm.
• Utilized the first two intrinsic mode functions (IMFs) to construct the SDC-L feature.
• Evaluated SDC-L using Support Vector Machine (SVM), k-Nearest Neighbor (k-NN), Random Forest (RF), Multilayer Perceptron (MLP), Convolutional Neural Network (cNN), and a cNN-LSTM hybrid model.
• Assessed classifier performance using accuracy, precision, recall, and F1-score.

Main Results:

• SDC-L effectively preserves temporal dependency information while reducing hyperspectral dimensionality.
• The first two IMFs were sufficient for constructing the SDC-L feature.
• SVM, MLP, and the cNN-LSTM hybrid model demonstrated superior performance with SDC-L.
• Other classifiers achieved comparable results across different features.

Conclusions:

• SDC-L is a promising novel feature for enhanced lung sound classification.
• The proposed method offers improved feature extraction for complex respiratory signals.
• SDC-L combined with advanced machine learning models can significantly improve diagnostic accuracy for respiratory conditions.