Classification and preliminary generalization of swallowing sounds measured by a low-cost contact microphone

Abstract

A simple and noninvasive acoustic approach for supporting swallowing assessment is needed. This study examined whether swallowing sounds recorded with a low-cost contact microphone contain discriminative acoustic information and show preliminary transferability to clinical data. Sounds from four healthy adult males were analyzed using STFT, Mel-spectrogram, and CWT features with SVM and k-NN classifiers. STFT- and Mel-based SVM models achieved test accuracies of 98.5% and 96.7%, respectively. In a preliminary application to clinical recordings, 3-cc and 5-cc thickened-water swallows were correctly recognized, whereas the 1-cc event was not. These results support the feasibility of low-cost acoustic swallowing assessment.