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

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning

Background:

• Dysphagia (swallowing disorder) necessitates continuous monitoring of pharyngeal and laryngeal functions.
• Current monitoring methods are limited to clinical settings, restricting long-term patient observation.
• Accurate, real-time data is crucial for effective dysphagia management.

Purpose of the Study:

• To develop and validate a soft, skin-attachable throat vibration sensor (STVS) for autonomous detection of swallowing events.
• To implement an ensemble-based deep learning model for accurate classification of throat-related events from sensor data.
• To assess the feasibility of a ubiquitous monitoring system for remote dysphagia management.

Main Methods:

• Development of a soft skin-attachable throat vibration sensor (STVS) to capture subtle throat vibrations.
• Utilizing an ensemble-based deep learning model integrating multiple neural networks for event classification.
• Training and testing the model on multi-modal acoustic features of throat-related events.
• Continuous data stream acquisition and automated event detection.

Main Results:

• The STVS demonstrated accurate recording of throat vibrations, including subtle swallowing sounds, with minimal noise interference.
• The deep learning model achieved a high classification accuracy of 95.96% for identifying throat-related events.
• The system successfully demonstrated autonomous detection and classification of events from continuous data.

Conclusions:

• The developed STVS and deep learning model offer a feasible solution for ubiquitous, non-invasive dysphagia monitoring.
• Wearable technology can significantly enhance patient monitoring and outcomes for swallowing disorders outside clinical environments.
• This approach holds promise for improving the management and quality of life for individuals with dysphagia.