2023-06-19 21:16:13
Why is this important?
Respiratory disorders related to the various pulmonary pathologies are detectable by stethoscope but the ability to make a differential etiological diagnosis by simply listening to them remains limited for the human ear. This limitation can be overcome thanks to automatic learning (deep learning), which has the potential to finely discriminate between almost similar audio signals. More specifically, convolutional neural networks, which have been developed in image recognition by machine learning, have more recently been adapted to the field of auditory signals.
In this approach, the diagnostic prediction is made a priori from recordings acquired at several anatomical locations in each patient. In addition, the algorithm initially developed on data collected independently was validated via an external cohort in order to ensure the interpretability and predictive robustness of the model, which might not be achieved for the tools previously developed. To this end, this study included a cohort of patients whose diagnoses, age and geographical distribution were varied.
Methodology
An international multicentre observational cohort (5 participating countries) was set up by recruiting 572 patients under 16 years of age admitted to an outpatient setting in a pediatric hospital department made up of cases (71%) with pneumonia, bronchiolitis or respiratory wheezing (asthma or obstructive bronchitis) or controls (29%). Those with known chronic respiratory or heart disease were excluded. Each was the subject of digital recordings (average duration 28.4 seconds) of their breathing on 8 anatomical sites via an electronic stethoscope. Recordings were categorized into one of four clinical diagnostic categories (control, pneumonia, wheezing, and bronchiolitis) following comprehensive medical evaluation by an expert pediatrician. The final model also incorporated several clinical variables (age, respiratory rate). The two most numerous cohorts among the 5 countries were used for the development phase of the algorithm and those recruited by the other three were used for the external validation.
DeepBreath has been trained to binary discriminate each type of diagnosis once morest the other three considered as a whole. Predictions were obtained for each patient by combining predictions from each of the 8 anatomical sites.
Principle results
DeepBreath is able to discriminate healthy patients from sick patients with an internal area under the curve (AUROC) of 0.931 and 0.887, according to external validation. The internal and external AUROC values associated with the predictive ability of pneumonia were 0.75 and 0.89 respectively, as well as 0.91 and 0.74 for wheezing and 0.94 and 0.87 for bronchiolitis.
DeepBreath performance was maintained when the number of recordings and their duration were reduced to 4 anatomical positions and 5-10 seconds respectively.
1687215057
#DeepBreath #simple #electronic #stethoscope #making #pediatric #diagnoses
