Deep Learning-Based Swallowing Monitor for Realtime Detection of Swallow Duration.

Abstract

Aspiration pneumonia is a life-threatening disease for the elderly. To prevent its risk, regular swallowing assessment is necessary; however, current screening tools for swallow assessment are not widely available and medical experts are insufficient. As a portable assessment tool, we have been developing a smartphone-based realtime monitoring device (GOKURI) which can evaluate swallowing ability based on swallow sounds. For better detection accuracy of the system, we integrated a deep learning model which was developed based on the swallowing anatomy. In this paper, we provide a detailed analysis to see how the swallow sounds detected by the deep learning-based monitor correspond to the actual swallow activities. Also, as an example of practical application of the system, we analyzed the changes of the swallow abilities over time by recording swallow sounds twice for the same participants at a nursing home. To minimize the risk of aspiration pneumonia, caregivers need to understand the disability levels of the patient's swallows so that safe feeding assistance can be provided. The result of this paper implies the possibility of using GOKURI as a daily swallowing monitor with minimum interventions.