Deep Learning-based Artificial Intelligence in Audio based Analysis of Swallowing using Cervical Auscultation

Abstract Swallowing problems (dysphagia) is associated with significant morbidity and mortality therefore diagnosis and treatment of dysphagia is important. Diagnostic tests include screening procedures, clinical swallowing examinations, and instrumental examination procedures. A non-invasive diagnostic option is auscultation of the swallowing act. However, there are different statements about the reliability and validity of the manual execution of this test. We developed a mobile hardware system to record cervical sounds using two microphones on the neck to acquire audio a data set. To generate ground truth data, fiberendoscopic swallow examinations were performed simultaneously to identify dysphagia. In order to diagnostically assess the swallowing sounds a spectrogram based classification pipeline was developed. In a first step this enabled us to identify different swallowing patterns in healthy individuals. With an accuracy of ~95%, we were able to reliably detect swallows within audio recordings, while the classification of types of swallow (dry, water, solid food) indicate the need for further improvements within the project ahead. In the future, we anticipate AI based analysis of auscultated swallowing sounds to detect swallowing disorders and aspirations.