Multimodal deep ensemble classification system with wearable vibration sensor for detecting throat-related events

Abstract

Dysphagia, a swallowing disorder, requires continuous monitoring of throat-related events to obtain comprehensive insights into the patient's pharyngeal and laryngeal functions. However, conventional assessments were performed by medical professionals in clinical settings, limiting persistent monitoring. We demonstrate feasibility of a ubiquitous monitoring system for autonomously detecting throat-related events utilizing a soft skin-attachable throat vibration sensor (STVS). The STVS accurately records throat vibrations without interference from surrounding noise, enabling measurement of subtle sounds such as swallowing. Out of the continuous data stream, we automatically classify events of interest using an ensemble-based deep learning model. The proposed model integrates multiple deep neural networks based on multi-modal acoustic features of throat-related events to enhance robustness and accuracy of classification. The performance of our model outperforms previous studies with a classification accuracy of 95.96%. These results show the potential of wearable solutions for improving dysphagia management and patient outcomes outside of clinical environments.