Development of First prototyped Algorithm for early detection for aspiration risk in subjects with swallowing

Year Submitted: 2025
Published Date: 08 April 2025

Tags: Automation, Information ＆ Communication Technology, Robotics, Care Process ＆ Redesign, Workflow Redesign, Productivity, Cost Saving

About this Content

Aims

The aim of the project is to achieve proof of concept (POC) in developing a prototype algorithm, which detects aspiration risk among subjects with swallowing impairment. Considering the costing, feasibility and simplicity, an acoustic analysis of swallowing sounds might be useful as a non-invasive clinical procedure to document and differentiate dysphagia symptomatology from normal. Various types of sensors have been used for swallowing detection purposes including accelerometers (S. Damouras et al., 2010), electromyography (EMG) sensors (Amft and Tröster, 2006) a miniature hearing-aid microphone (Lear, 1965), contact microphone (Mackowiak et al., 1967), miniature dynamic earphone (Cunningham and Basmajian, 1969), throat microphone (Selley, 1990), piezo-electric type pick-up (Ertekin, 2002), dynamic microphone (Lebel et al., 1990), and electret condenser microphone (Murti et al.,1980). Although acoustic characteristics were extracted from the swallowing sounds acquired through these devices, no empirical evidence associated with the accurate effectiveness of the detector unit in free-living conditions was reported to the best of our knowledge.
The algorithm developed aims to be effective in retrieving correct events and omitting non-swallow events while maintaining low processing effort for classifying aspiration signs and the swallowing disorders. The developed algorithm, via signal processing using the swallowing sound waves, aims to accurately detect and determine the presence of dysphagia and the associated aspiration risk. For this initial proof-of-concept stage, the algorithm developed will serve to differentiate between normal swallowing and abnormal swallowing sound and detect aspiration.

Methods

We recruited all patients who were recommended for Fiberoptic Endoscopic Evaluation of Swallowing (FEES) procedure by their attending speech therapists. FEES is a type of objective test procedure that uses a fiberoptic nasoendoscope placed in the throat to obtain the actual visualization of the pharynx to determine the effectiveness of the patient’s swallow. Patients were fed various consistencies of fluids and food based on patients’ swallowing impairment and risk of aspiration. During the procedure, a 3M Littman electronic stethoscope was placed around the patient’s lateral neck for the audio recording. To minimize the discomfort and enhance the clarity of sound collection, patients with injury or complication at the neck area were not eligible in the study. For the purpose of this study, only trials of nectar fluid consistency were recorded. These recordings were then analyzed and annotated using aspiration-penetration scale (PAS) by speech therapist. The annotated data was further analyzed to develop an algorithm using imaging processing of the swallowing sound wave with convolution neural network in different models for the differentiation between swallowing and non-swallowing sounds. We defined swallowing sound as “normal” swallowing sounds when they occurred without aspiration and penetration while “non-swallowing” sounds mainly refers to aspiration signs i.e. cough or throat clearing which occurred during the procedure. The annotated data was sent to our collaborator, NUS master students from the Institute of System Science to develop the algorithm.

Results

33 male and 14 female patients were enrolled in the study. The annotated swallowing sounds were analyzed using VGG1, VGG19 and Inception V3 models. Initial run of the model using 30 epochs highly affected the accuracy of the overall validity of the result. This was improved with use the of 10 epochs.

Analysis using VGG16 model showed 68% accuracy with 0.20 precision and recall scoring. With the VGG19 model, it achieved 74% accuracy, 0.08 precision and 0.1 recall scoring. Interesting in the InceptionV3 model, 63% accuracy, 0.47 precision and 0.6 recall scoring were obtained.

From this study, it shows that the algorithm developed is able to differentiate between “normal” and “non-swallowing” sounds. While the level of accuracy is considerably low as a predictive algorithm, all models presented, especially the InceptionV3, shows the ability to classify sound waves and demonstrates the potential to improve on the accuracy of the algorithm by increasing sample size. The possibility of taking this algorithm to the next level in classifying sound waves with more annotations will serve our eventual goal of developing a tool which is accurate and easily-accessible to frontline staff, and is able to identify people with dysphagia early for appropriate intervention by ST and/or prevention measures to be put in place while awaiting ST service.

Lessons Learnt

More work needs to be done in fine-tuning the algorithm by increasing the sample size to improve its accuracy and precision, as well as the use of a better recording tool to improve the quality of sounds recorded

Keywords

Technology, Screening Device, Algorithm Development, Dysphagia Screening Tool, Aspiration Risk

Innovators' Details

Connect with this contributor!

T’ng Kuan Chen Zenne - Kuan_chen_tng@ttsh.com.sg

Project Attachment