Work

Automatic Segmentation and Identification of Pharyngolaryngeal Activities in Swallis DSA Signals

Year: 2022
Type: conference-paper
Abstract: Introduction - High-resolution cervical auscultation offers a promising alternative to swallowing screening by analysing vibroacoustic signals captured on the throat. This study evaluates a new method to automatically segment and identify pharyngolaryngeal events of interest in signals recorded from healthy volunteers with the Swallis DSA medical device.Material and Methods - The dataset used for this study is part of an ongoing data collection and is composed of 243 swallows, 169 phonations, and 19 expectorations by 6 different subjects. The predefined scenario includes swallows of different volumes and textures as water or biscuit. The signals are first segmented using an activity detection recurrent neural network and then classified into 4 different labels (swallowing, phonation, expectoration, "other event") by a convolutional neural network. Leave-one-out technique was used so that no record of a tested subject is used on training. The proportion of well classified segments on the total number activity segments detected is used as performance score.Results and Discussion - The best score obtained so far is 92% across the 4 classes. However, some events have not been correctly identified, especially some spontaneous saliva swallows and consecutive swallows occurring while drinking a full glass of water, which present specific patterns in the captured signals. Increasing the database should lead to more accurate identification of these events. Conclusion - The method applied to the dataset provides good overall segmentation of events of interest. The combination of Swallis DSA and Swallis-Algorithms could help practitioners to monitor the swallowing activity of their patients over variable lasting observations. The non-invasive aspect of the exam might allow to integrate it to a routine screening and thus improve the care of dysphagia condition, a public health issue of difficult management.
Source: HAL
Language: English
Open Access status: green
OpenAlex ID: W4388729575
FWCI: 0.658
Cites:
Cited by: 1
Related to: 10