Method for Classifying Spectrograms of Acoustic Signals from Swallowing of Individuals with and without Dysphagia Using a Convolutional Neural Network.

Abstract

Dysphagia is highly prevalent and may lead to important clinical complications. Although traditional diagnostic methods are effective, they have limitations, including high cost and patient discomfort. Cervical auscultation is a non-invasive technique for monitoring swallowing sounds and may support dysphagia assessment. To present a method for analyzing and classifying acoustic signals obtained through cervical auscultation of individuals with and without dysphagia using a machine learning approach. Cross-sectional study including swallowing assessments from individuals aged 18 years or older. Simultaneously with the clinical assessment of dysphagia, swallowing sounds were recorded using an Eko Core amplifier coupled to a 3 M-LITTMANN stethoscope. The stethoscope was positioned over the lateral edge of the trachea just below the participant's cricoid cartilage, and the participant was instructed not to speak or make unnecessary sounds during the recording to minimize background noise. Acoustic signals from individuals with and without dysphagia were captured with a digital stethoscope and stored for analysis. A total of 178 swallowing signals were selected, segmented, and augmented to 1,888 isolated events, which were converted into spectrograms using the Short-Time Fourier Transform. The learning model used was a convolutional neural network associated with a cross-validation system. The model achieved an accuracy of 77.7% and a sensitivity of 76.6%, but relatively low precision (64.1%). The area under the ROC curve was 0.823. This study advances the automation of dysphagia assessment by showing that acoustic analysis supported by neural networks may complement traditional methods. Future investigations should focus on improving model robustness and clinical applicability to support a more accessible and less invasive diagnostic approach.