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

Abstract

Background: Swallow kinematic analysis is the process of evaluating the physiological attributes of swallowing. Although videofluoroscopic swallowing studies (VFSS) are the gold standard, non-invasive and radiation-free high-resolution cervical auscultation (HRCA) has been extensively studied as a potential alternative. HRCA has been effectively employed for individual kinematic analysis tasks, but its application to conduct a comprehensive multitask analysis has not yet been explored.

Methods: In this study, we use shared parameter multi-task learning with transformer encoders to develop an integrated swallow kinematic analysis framework using sensory information presented in HRCA signals. We used HRCA signals recorded from 120 patients and labeled by speech-language pathologists to train and evaluate the framework. This framework analyzes multiple swallowing kinematics landmarks from individual swallows within HRCA signals.

Results: The proposed framework achieved 92% accuracy, 89% sensitivity, and 92% specificity in a 10-fold cross-validation procedure over the main dataset, and maintained over 85% accuracy when tested on a never-seen independent dataset collected from a different cohort of subjects. These results not only demonstrate strong generalization capabilities but also outperform all state-of-the-art and baseline models reported in the literature for individual task detection in swallowing assessment.

Conclusion: The results indicate that the multi-task framework provides a detailed and objective analysis of key components of swallowing such as upper esophageal sphincter opening duration and distension, laryngeal vestibule closure duration, and hyoid bone displacement with comparable performance to single-task detection algorithms and VFSS-based human ratings. Finally, we presume that the comprehensive swallowing parameter analysis would help speech-language pathologists in the process of decision making and diagnosis of swallowing impairments.

Keywords: High-resolution cervical auscultations; Multi-task learning; Swallowing kinematics; Transformers.