High-Accuracy Lung Sound Classification for Healthy versus Unhealthy Diagnosis Using Artificial Neural Network High-Accuracy Lung Sound Classification for Pulmonary Disease Diagnosis Using Artificial Neural Network

Weiwei Zhang¹Xinyu Li²Qiao Liu³Xiangyang Zheng⁴Yisu Ge⁵Xiaotian Pan^6*Yu Zhou¹

• ¹Third Affiliated Hospital of Wenzhou Medical University, Wenzhou, Zhejiang Province, China
• ²China Telecom Corporation Limited Zhejiang Branch, Hangzhou, China
• ³Wenzhou Medical University, Wenzhou, Zhejiang Province, China
• ⁴Wenzhou Institute of Technology, Wenzhou, Zhejiang Province, China
• ⁵Wenzhou University, Wenzhou, China
• ⁶Hangzhou Dianzi University, Hangzhou, China

This study presents a comprehensive approach to classify lung sounds into healthy and unhealthy categories using a dataset collected from 112 subjects, comprising 35 healthy individuals and 77 patients with various pulmonary conditions, such as asthma, heart failure, pneumonia, bronchitis, pleural effusion, lung fibrosis, and chronic obstructive pulmonary disease (COPD), grouped as unhealthy. The dataset was obtained using a 3M Littmann® Electronic Stethoscope Model 3200, employing three types of filters (Bell, Diaphragm, and Extended) to capture sounds across different frequency ranges. We extracted five key audio features-Spectral Centroid, Power, Energy, Zero Crossing Rate, and Mel-Frequency Cepstral Coefficients (MFCCs)-from each recording to form a feature matrix. A Multi-Layer Perceptron (MLP) neural network was trained for binary classification, achieving accuracies of 98%, 100%, and 94% on the training, validation, and testing sets, respectively. This partitioning ensured the model's robustness and accuracy. The high classification accuracy achieved by the MLP neural network suggests that this approach is a valuable decision-support tool for identifying healthy versus unhealthy lung sounds in clinical settings, facilitating early intervention while maintaining computational efficiency for offline implementation. This study presents a comprehensive approach to classify lung sounds using a dataset collected from 112 subjects, comprising both healthy individuals and patients with various pulmonary conditions such as asthma, heart failure, pneumonia, bronchitis, pleural effusion, lung fibrosis, and chronic obstructive pulmonary disease (COPD). The dataset was obtained using a 3M Littmann® Electronic Stethoscope Model 3200, employing three types of filters (Bell, Diaphragm, and Extended) to capture sounds across different frequency ranges. We extracted five key audio features-Spectral Centroid, Power, Energy, Zero Crossing Rate, and Mel-Frequency Cepstral Coefficients (MFCCs)-from each recording to form a feature matrix. A Multi-Layer Perceptron (MLP) neural network was trained on the extracted features. The dataset was split into training, validation, and testing sets, achieving classification accuracies of 98%, 100%, and 94% respectively.