An innovative AI algorithm created by researchers at EPFL and University Hospital Geneva shows promise for improving diagnosis and treatment of respiratory diseases. The new study, published in Nature Digital Medicine, details the DeepBreath algorithm's ability to interpret auscultation sounds and detect pneumonia, asthma, and bronchiolitis.
Developed by Dr. Mary-Ann Hartley's team at EPFL, DeepBreath was trained on audio recordings from nearly 600 pediatric outpatients in Switzerland, Brazil, Senegal, Cameroon, and Morocco. "The diversity and careful collection of respiratory sounds make this study stand out," said Dr. Hartley. Respiratory diseases are the leading preventable cause of mortality in this age group."
DeepBreath demonstrated strong performance when tested on data from different countries, indicating its potential for generalization. "There are many differences between hospitals globally that could affect sounds, like background noise and stethoscope use," explained Dr. Hartley. "But with enough data, the algorithm seems robust to these variations."
Analyzing the algorithm’s predictions revealed it truly relies on breathing sounds, not confounding factors like background noise. This helps validate the interpretability of DeepBreath's decisions.
The startup Onescope, founded by Professor Alain Gervais of University Hospital Geneva, will launch a “smart stethoscope” integrating the DeepBreath technology. "Non-consumable diagnostic tools like this intelligent stethoscope offer guaranteed stability and potential for continuous improvement as more data is collected," said Dr. Hartley.
The research team is preparing DeepBreath for real-world use. Next steps include testing it on more patients using the new Pneumoscope stethoscope, which also records temperature and blood oxygen saturation. "Combining these additional signals will likely further boost accuracy," said Dr. Hartley.
The study demonstrates the power of AI to improve respiratory disease diagnosis globally. DeepBreath offers clinicians an innovative tool to enhance treatment and outcomes.
