A new artificial intelligence algorithm developed at EPFL and University Hospital Geneva (HUG) will be used in the intelligent Pneumoscope stethoscope with the potential to improve the treatment of respiratory diseases.

When air passes through a maze of small passages in the lungs, it emits a characteristic whistling sound. When these passages narrow in asthmatic inflammation or are clogged with infectious secretions in bronchitis, the sound changes in a characteristic way. The detection of these diagnostic signs using a stethoscope is an unavoidable element of almost every medical examination.

However, despite two centuries of experience in using stethoscopes, the interpretation of auscultation is still extremely subjective. At the same time, depending on where in the world you are, the same sound can be described in different ways. The accuracy is also influenced by the experience of the medical worker and his specialization.

These difficulties make the task ideal for deep learning, which is able to distinguish sound patterns more objectively. The technology has already shown that it can complement human perception when interpreting a number of complex medical examinations, such as X-rays and MRI.

Now, in a new study published in the journal Nature Digital Medicine, the EPFL research group has described its DeepBreath AI algorithm, which demonstrates the potential of automated interpretation in the diagnosis of respiratory diseases.

"What makes this study particularly unique is the diversity and careful collection of a bank of auscultative sounds," said senior study author Dr. Mary-Ann Hartley. - Almost 600 pediatric outpatient patients were recruited in five countries - Switzerland, Brazil, Senegal, Cameroon and Morocco. Breathing sounds were recorded in patients under the age of fifteen with the three most common types of respiratory diseases - radiographically confirmed pneumonia and clinically diagnosed bronchiolitis, as well as asthma."

"Respiratory diseases are the number one cause of preventable mortality in this age group," explains Professor Alain Gervais, head of the Department of Pediatric Medicine at HUG and founder of the startup Onescope, which will launch a stethoscope with the DeepBreath algorithm. - This work is an excellent example of successful cooperation between HUG and EPFL, between clinical research and basic science. The pneumoscope with the DeepBreath algorithm is a breakthrough innovation for the diagnosis and treatment of respiratory diseases."

Dr. Hartley's team is leading the development of AI for Onescope. "Reusable, non-consumable diagnostic tools, such as this intelligent stethoscope, have the unique advantage of guaranteed stability," she explained. "AI tools also have the potential for continuous self-improvement, and I hope that we will be able to extend the algorithm to other respiratory diseases and population groups by obtaining additional data."

The DeepBreath algorithm was trained on patients from Switzerland and Brazil, and then tested on records from Senegal, Cameroon and Morocco, which gives an idea of the geographical generalizability of the tool. "You can imagine that there are many differences between emergency departments in Switzerland, Cameroon and Senegal," says Dr. Hartley and lists examples of "the soundscape of background noise, the way a doctor holds a stethoscope, a sound recording device, epidemiology and local diagnostic protocols."

With a sufficient amount of data, the algorithm should be resistant to these nuances and find a signal among the noise. Despite the small number of patients, DeepBreath has demonstrated impressive results in different locations, indicating the potential for further improvement with an increase in the amount of data.

A particularly unique contribution of the research was the inclusion of methods aimed at revealing the inner workings of the algorithm's "black box". The authors were able to demonstrate that the model really uses the respiratory cycle for its predictions, and show which parts of it are most important. The proof that the algorithm actually uses breathing sounds, and does not "cheat" by using biased signatures in background noise, is a critical gap in the existing literature.

The interdisciplinary team is working on preparing the algorithm for real use in the intelligent stethoscope Pneumoscope. The next important task is to repeat the study on more patients using recordings from this newly developed stethoscope, which also records the patient's temperature and blood oxygen saturation. "Combining these signals together is likely to further improve forecasts," predicts Dr. Hartley.