Design and Development of an Intelligent System Applied to the Interpretation of Lung Auscultation

Abstract

Auscultation is a widely used technique in clinical practice for diagnosing and monitoring respiratory diseases; however, its effectiveness is limited by subjectivity, dependence on specialized expertise and environmental factors. To overcome these limitations, this work presents the design, development, and proof of concept of an intelligent decision support system for the interpretation of lung auscultation. The proposed system generates a multimodal representation that integrates acoustic data from auscultation—transformed into a time-frequency representation using the Continuous Wavelet Transform with the Generalized Morse Wavelet family—with clinical information, including body mass index and the presence of respiratory diseases, such as chronic obstructive pulmonary disease, asthma, interstitial lung disease, bronchiectasis and emphysema. The resulting image is processed by two independent convolutional neural networks based on the SqueezeNet architecture, which estimate the likelihood of pathological lung sounds in the left and right lungs separately. This approach enables precise localization of abnormalities and supports a more objective assessment of auscultation findings. Evaluation on a reserved test set yielded an area under the ROC curve (AUC) of 0.77 for both classifiers. Beyond its immediate diagnostic utility, this study introduces a novel multimodal data-fusion framework that combines clinical and acoustic information for automated lung sound interpretation. Despite the promising results, it should be noted that the system is still in a conceptual phase of development, so further validation will be needed before its implementation in clinical practice.