Machine learning-driven lung sound analysis: novel methodology for asthma diagnosis
Bildiriler > Özet Bildiri
Yazarlar: Ihsan Topaloğlu, Gulfem Ozduygu, Cagri Atasoy, Guntug Batihan, Damla Serce, Mutlu Onur Gucsav, Arif Metehan Yıldız, Turker Tuncer, Sengul Dogan, Prabal Datta Barua
M-Health/e-health, , , Uluslararası, Özet Bildiri, ss PA4972-PA4972
DOI: 10.1183/13993003.congress-2025.pa4972
Özet
Background: Asthma is a chronic airway inflammatory disease with variable airflow limitation and intermittent symptoms. In well-controlled asthma, auscultation and spirometry often appear normal. Moreover, bronchial provocation tests are limited by potential risk of inducing acute bronchoconstriction. This study aimed to develop a noninvasive, objective and reproducible diagnostic method using machine learning-based lung sound analysis for the early detection of asthma, even during stable periods. Methods: We developed a machine learning algorithm to classify 60 controlled asthma and 60 healthy individuals using digital stethoscope recordings. Controlled asthma was defined according to GINA criteria and was supported by normal spirometry, the absence of pathological auscultation findings and exacerbations in the last three months. A total of 3600 respiratory sound samples were analysed, with feature extraction performed using Mel-Frequency Cepstral Coefficients and Tunable Q-Factor Wavelet Transform. Results: Pulmonary function tests differed significantly between groups, though asthma values remained within normal ranges. The Quadratic SVM and Narrow Neural Network models achieved accuracies of 99.86% and 99.63%. Quadratic SVM correctly classified 99.44% of controls and 99.89% of asthma, while the Narrow Neural Network achieved 99.11% and 99.83%. Both models had sensitivity, specificity and F1-scores above 99%. Conclusion: The machine learning algorithm has the potential to diagnose asthma with high accuracy even in patients with normal spirometry and clinical findings, facilitating early and reliable diagnosis.