Multi-channel Lung Sound Analysis for COPD Diagnosis Using Statistical Test and Artificial Neural Network

Chronic obstructive pulmonary disease (COPD), known as the most prevalent chronic lung disease, is associated with high global morbidity and mortality rates. This study included 21 participants with lung sounds recorded from their 16 posterior chest locations. The study was conducted in two phases: In the initial phase, statistical tests were employed to analyze the normalized energy of various frequency sub-bands within lung sounds to determine the frequency ranges that best differentiate healthy and COPD groups. Meanwhile, the second phase aimed to assess the diagnostic capabilities of four distinct lung regions and the 16 recording channels using a Multilayer Perceptron (MLP) neural network. To achieve this, lung sounds were divided into four groups, each representing a specific lung region. Mel-Frequency Cepstral Coefficients (MFCCs) were then extracted from these groups, and the performance of both the four lung regions and the initial 16 channels was evaluated using the MLP. Additionally, inspiratory and expiratory phases, as well as complete respiratory cycles, of lung sounds were independently analyzed to determine their discriminative abilities in both study phases. Statistical analysis revealed significant differences between groups across most channels within specific frequency ranges: 200–300 Hz during full respiratory cycle, 200–300 Hz and 500–600 Hz during inspiration, and 600–1000 Hz during expiration. Furthermore, according to the neural network analysis, the upper right lung region and channel 9 demonstrated superior performance, achieving accuracies of 94.3% for complete respiratory cycles and 97.99% for expiration, respectively.