Swallowing diffculties, commonly referred to as dysphagia, affect thousands of Americans every year. They have a multitude of causes, but in general they are known to increase the risk of aspiration when swallowing in addition to other physiological effects. Cervical auscultation has been recently applied to detect such diffculties non-invasively and various techniques for analysis and processing of the recorded signals have been proposed. We attempted to further this research in three key areas. First, we characterized swallows with regards to a multitude of time, frequency, and time-frequency features while paying special attention to the differences between swallows from healthy adults and safe dysphagic swallows as well as safe and unsafe dysphagic swallows. Second, we attempted to utilize deep belief networks in order to classify these states automatically and without the aid of a concurrent video uoroscopic examination. Finally, we sought to improve some of the signal processing techniques used in this field. We both implemented the DBSCAN algorithm to better segment our physiological signals as well as applied the matched complex wavelet transform to cervical auscultation data in order to improve its quality for mathematical analysis.;Keywords: cervical auscultation, dysphagia, deep learning, signal analysis, signal features, classification.
展开▼
