Non-invasive integrated swallowing kinematic analysis framework leveraging transformer-based multi-task neural...

Area of Science:

• Biomedical Engineering
• Clinical Diagnostics
• Speech-Language Pathology

Background:

• Swallow kinematic analysis evaluates physiological swallowing attributes.
• Videofluoroscopic swallowing studies (VFSS) are the gold standard but involve radiation.
• High-resolution cervical auscultation (HRCA) is a radiation-free alternative, but comprehensive multitask analysis is unexplored.

Purpose of the Study:

• To develop an integrated framework for comprehensive swallow kinematic analysis using HRCA signals.
• To apply shared parameter multi-task learning with transformer encoders for this analysis.
• To explore the application of HRCA for multitask swallowing assessment.

Main Methods:

• Utilized shared parameter multi-task learning with transformer encoders.
• Trained and evaluated the framework using HRCA signals from 120 patients, labeled by speech-language pathologists.
• The framework analyzes multiple swallowing kinematics landmarks from individual swallows within HRCA signals.

Main Results:

• Achieved 92% accuracy, 89% sensitivity, and 92% specificity in 10-fold cross-validation.
• Maintained over 85% accuracy on an independent dataset, demonstrating strong generalization.
• Outperformed state-of-the-art and baseline models for individual task detection in swallowing assessment.

Conclusions:

• The multi-task framework provides detailed, objective analysis of key swallowing components (e.g., UES opening, laryngeal vestibule closure, hyoid displacement).
• Performance is comparable to single-task algorithms and VFSS-based human ratings.
• The comprehensive analysis aids speech-language pathologists in decision-making and diagnosing swallowing impairments.