Deep Learning for Classification of Normal Swallows in Adults

Area of Science:

• Biomedical Engineering
• Clinical Diagnostics
• Machine Learning in Healthcare

Background:

• Cervical auscultation is a non-invasive technique for evaluating swallowing function.
• The diagnostic utility of cervical auscultation as a clinical classification tool requires further investigation.
• Deep Belief networks offer a promising approach for analyzing complex physiological signals.

Purpose of the Study:

• To evaluate the efficacy of Deep Belief networks in classifying swallowing events.
• To differentiate swallows from healthy individuals versus those with dysphagia using dual-axis vibration analysis.
• To determine the impact of analyzing single versus dual-axis vibration signals on classification performance.

Main Methods:

• Recorded dual-axis swallowing vibrations from 108 participants (55 healthy, 53 unhealthy).
• Utilized 1946 discrete swallow events for analysis.
• Applied Fourier transforms to vibration signals as input for single and multi-layer Deep Belief networks.

Main Results:

• Single and multi-layer Deep Belief networks showed similar performance when analyzing a single vibration signal.
• Multi-layered Deep Belief networks achieved 5-10% higher accuracy and sensitivity when analyzing both dual-axis signals concurrently.
• Concurrent analysis of dual-axis vibrations highlights the importance of higher-order signal relationships.

Conclusions:

• Deep Belief networks are effective for classifying swallowing performance based on cervical auscultation.
• Analyzing concurrent dual-axis swallowing vibrations enhances classification accuracy, suggesting clinical utility.
• This approach holds potential for improved objective assessment of swallowing disorders.