300 Section - Multi-Scale Neural Framework for Objective sEMG-Based Swallowing Impairment Assessment

Objectives : Swallowing assessment is fundamental in identifying early neuromuscular abnormalities and preventing aspiration-related complications. Although surface electromyography (sEMG) has been widely used for monitoring suprahyoid muscle activity, few studies have quantitatively evaluated how specific bolus characteristics influence neuromuscular patterns. This study aims to (1) quantify sEMG signal differences elicited by swallowing tasks with varying bolus volumes and textures, and (2) develop a deep-learning–based classification framework capable of accurately identifying swallowing conditions. This approach provides a foundation for objective, non-invasive screening tools for dysphagia risk evaluation.

Design: We conducted a controlled experimental study involving thirty healthy volunteers who performed structured swallowing tasks with three bolus volumes (3 mL, 5 mL, 8 mL) and five IDDSI-defined textures (0–4). sEMG recordings were obtained from different regions using standardized electrode placement for comparative analysis. A new multi-branch deep-learning model integrating multi-scale convolution and attention-based feature fusion was developed to classify swallowing conditions.

Results: Quantitative analyses revealed systematic and significant differences in sEMG amplitude, activation duration, and signal complexity across both bolus volume and texture conditions. Larger volumes generated higher peak activation and prolonged elevation phases, while thicker textures elicited increased waveform complexity and entropy. The proposed deep-learning framework demonstrated strong discriminative capability, achieving balanced improvements in AUROC, accuracy, precision, recall, and F1 score compared with traditional machine-learning classifiers. Ablation experiments verified the contributions of multi-scale feature extraction and the attention-based fusion mechanism.

Conclusion: This study provides a comprehensive quantitative characterization of how bolus properties modulate swallowing-related sEMG patterns and introduces a high-performing deep-learning model capable of accurately classifying swallowing tasks. The findings highlight the feasibility of non-invasive, EMG-based digital biomarkers for swallowing assessment and offer a technical foundation for future clinical applications in dysphagia screening, rehabilitation monitoring, and intelligent swallowing-function evaluation systems.