SafeSwallow: A Compact Closed-Loop Wearable Neuroprosthesis for Multiphase Dysphagia—Motivation, Architecture, Work Plan, Algorithms, and Dual-Track Performance Evaluation
SafeSwallow: A Compact Closed-Loop Wearable Neuroprosthesis for Multiphase Dysphagia—Motivation, Architecture, Work Plan, Algorithms, and Dual-Track Performance Evaluation
Where did the research take place?
The study site has not been established. Author addresses may differ from where the research occurred.
Publication status: preprint
A plain-language reading has not been prepared for this paper yet.
Original abstract
Dysphagia comprises aspiration risk as well as oral-phase impairment, post-swallow residue, reduced sensation, delayed pharyngeal onset, and incomplete hyolaryngeal elevation. Current daily-life approaches generally separate wearable sensing from electrical stimulation and rarely provide per-swallow, safety-bounded adaptation. This paper presents SafeSwallow™, a compact wearable neuroprosthesis that integrates around-ear or glasses electroencephalography (EEG) for anticipatory arming, submental surface electromyography (sEMG), laryngeal bioimpedance (BI), contact acoustics, and inertial measurement for swallow verification and elevation feedback. The controller combines reactive multimodal gating, anticipatory onset prediction, explicit electromechanical-delay (EMD) compensation, elevation-state estimation, and iterative learning control (ILC) of stimulation dose under a non-learned Tier-0 safety supervisor. The study evaluates the architecture on two complementary tracks: a physiology-inspired 1-kHz synthetic closed-loop experiment with healthy and Parkinsonian phenotypes and a public videofluoroscopy-aligned high-resolution cervical auscultation (HRCA) dataset. On held-out synthetic trials, closed-loop control achieved 100.0% Q1 force success with 0.0% misses and 0.0% false stimulation, while reducing normalized elevation error from 0.253 under matched open-loop dosing to 0.116. On the public HRCA test set, the improved causal front-end increased detection from 83.1% to 94.9% and modeled Q1 force success from 59.2% to 92.9%. These results support engineering feasibility of the proposed timing and control architecture, while demonstrating that clinical translation requires paired multimodal datasets, measured electrode-specific EMD, and prospective human validation.