Pediatric Pneumonia Diagnosis: Integration of a Self-Assembled Digital Stethoscope with Raspberry Pi and 1D CNN Model

Globally, pneumonia is the leading infectious killer of children under five, accounting for 14% of all deaths. Approximately, three children die every hour from pneumonia in Bangladesh, which accounts for 24% of all childhood deaths- exceeding the global average of 24,000. These deaths are preventable with early detection of pneumonia, which is commonly done by radiologic imaging but is often not accessible to remote areas. Another diagnostic technique is through auscultation of lungs using a stethoscope, although it is subject to human errors, environmental noise and can be misdiagnosed. Hence, in our work a self-assembled digital stethoscope integrated with Raspberry-Pi, a pocket-size computer is used to analyze and predict pneumonia. In the study, a database was developed from 10 pneumonia-infected and 9 healthy subjects, collecting a total of 72 samples from multiple chest locations. The dataset was preprocessed and augmented through methods like noise addition, time stretching, and pitch shifting to simulate varied real-world scenarios. A 1D-Convolutional Neural Network (CNN) model was trained to identify patterns in the lung sounds indicative of pneumonia. The audio data was represented as feature vectors, including Mel-Frequency Cepstral Coefficients (MFCCs) and Zero Crossing Rate (ZCR). The CNN model comprised convolutional, pooling, and fully connected layers, with dropout layers to prevent overfitting and SoftMax activation in the output layer for probabilistic predictions. After rigorous training and testing, the model achieved an accuracy of around 86%, demonstrating its potential in early detection of pneumonia.