Experimental classification of acoustic emissions from oral behaviors including bruxism using deep learning

Bruxism is a parafunctional oral behavior that affects a large percentage of the population. Images of sounds produced by jaw clenching, tooth grinding, reading, eating and drinking by 18 participants were recorded from inside the ear and classified using ResNet-50. Data were segmented into 1 s windows with 50 % overlap. RGB images were used: time series assigned to red, Short Time Fourier Transformation (STFT) assigned to green, and the spectrogram assigned to blue. Three classifiers were examined: 2-Class (Grinding and Pause), 4-Class (Eating, Grinding, Pause, Reading), and 6-Class (Clenching, Grinding, Reading, Eating, Drinking, and Pause). Overall accuracy was 84.31 %, 72.79 %, and 50.97 % for 2-Class, 4-Class, and 6-Class respectively.