Normal and Abnormal Lung Sound Classification for an Embedded System Implementation

Breath sounds serve as critical indicators of respiratory health, reflecting air movement, changes in lung tissue, and secretions within the tracheobronchial tree. While auscultation is a common clinical practice for detecting abnormal respiratory sounds, it is inherently subjective and prone to errors. This paper presents a novel methodology for processing respiratory sounds using frequency analysis based on Fourier transform and power spectral density (PSD) with Welch method. It introduces three multilayer perceptron (MLP) architectures designed for the automatic classification of lung sounds into normal and abnormal categories. The proposed approach analyzes 4-second respiratory cycles, employing different window sizes to limit the frequency response output of the PSD method, aiming to reduce the complexity of MLP architectures while maintaining high accuracy. The models were trained and tested on two public datasets, ICBHI-2017 and HF_LUNG_V1, achieving accuracies of 100%, 95%, and 91% for MLP1, MLP2, and MLP3, respectively. These results confirm the effectiveness of this approach in automatically classifying normal and abnormal respiratory sounds using MLP architectures. Furthermore, due to their simplicity, future efforts will focus on deploying these systems on hardware platforms, contributing to the development of portable devices that assist medical staff in identifying abnormal lung sounds.