Evaluation of an automated swallow-detection algorithm

Written by Gabi Constantinescu

Gabi is an Adjunct Assistant Professor in the Department of Communication Sciences and Disorders at the University of Alberta. She is also the CPO for True Angle. Her doctoral work directly influenced the design and development of the Mobili-T®, a mobile therapy system for swallowing exercise.

April 11, 2020

Our team created a way for the Mobili-T® to detect when a patient’s swallowing muscles contract. This is known as an automated swallow-detection algorithm. This algorithm makes sure that signals arising from swallow or swallow-like exercises are reinforced, while non-swallow movements are ignored.


Mobile health (mHealth) technologies may offer an opportunity to address longstanding clinical challenges, such as access and adherence to swallowing therapy. Mobili-T® is an mHealth device that uses surface electromyography (sEMG) to provide biofeedback on submental muscles activity during exercise. An automated swallow-detection algorithm was developed for Mobili-T®. This study evaluated the performance of the swallow-detection algorithm. Ten healthy participants and 10 head and neck cancer (HNC) patients were fitted with the device. Signal was acquired during regular, effortful, and Mendelsohn maneuver saliva swallows, as well as lip presses, tongue, and head movements. Signals of interest were tagged during data acquisition and used to evaluate algorithm performance. Sensitivity and positive predictive values (PPV) were calculated for each participant. Saliva swallows were compared between HNC and controls in the four sEMG-based parameters used in the algorithm: duration, peak amplitude ratio, median frequency, and 15th percentile of the power spectrum density. In healthy participants, sensitivity and PPV were 92.3 and 83.9%, respectively. In HNC patients, sensitivity was 92.7% and PPV was 72.2%. In saliva swallows, HNC patients had longer event durations (U = 1925.5, p < 0.001), lower median frequency (U = 2674.0, p < 0.001), and lower 15th percentile of the power spectrum density [t(176.9) = 2.07, p < 0.001] than healthy participants. The automated swallow-detection algorithm performed well with healthy participants and retained a high sensitivity, but had lowered PPV with HNC patients. With respect to Mobili-T®, the algorithm will next be evaluated using the mHealth system.

Keywords: Deglutition; Deglutition disorders; Dysphagia; Head and neck cancer; Mobile health; Surface electromyography; Swallow recognition; Visual biofeedback.

You May Also Like…