Fraunhofer Institute for Toxicology and Experimental Medicine
Chronic respiratory diseases have become a major global health challenge and, according to a publication by the Institute for Health Metrics and Evaluation (IHME) from April 2023, represent the third leading cause of death worldwide. In 2019, they were responsible for approximately 4 million deaths, and one in twenty people worldwide was affected by a chronic respiratory disease. Auscultation, i.e., listening to lung and heart sounds using a stethoscope, is the most commonly applied diagnostic method worldwide and provides healthcare professionals with invaluable insights into the structure and function of both organ systems. However, studies have shown that the reliability of this method often depends on the physician’s experience and hearing ability, which can lead to subjective and error-prone interpretation of the sounds. Therefore, it is crucial to pursue innovative approaches that enhance our understanding of respiratory sounds and enable the development of new, more reliable diagnostic tools.
Fraunhofer ITEM was involved in the project “DigitaLung” (Digital auscultation system for the differential diagnosis of lung diseases using machine learning), which ran until autumn 2025. The project was funded by the Medical Technology Program of the Federal Ministry of Education and Research and carried out in collaboration with ERKA Kallmeyer Medizintechnik GmbH & Co. KG, Hannover Medical School, and Leibniz University Hannover. Within the project, researchers at Fraunhofer ITEM developed and optimized a user-centered web application to support expert analysis of respiratory sounds. The web app is designed as a standalone tool and addresses two key aspects of the research objectives: the segment-wise visualization of breathing cycles and the integrated use of unsupervised machine learning algorithms.
The results obtained demonstrate that the applied machine learning methods enable effective classification of respiratory diseases. Furthermore, by specifically adjusting model parameters – particularly the loss functions – a balanced trade-off between sensitivity and specificity was achieved. This also holds true across different subgroups, for example with regard to age and sex.
