# Experimental Analysis of Automatic Discrimination Performance Between Simulated Bruxism and Non‐Bruxism Under Conscious Conditions Using Electromyography and Machine Learning

## TL;DR

This study explores using machine learning and EMG data to distinguish simulated bruxism movements with and without tooth contact from non-bruxism movements.

## Contribution

A novel machine learning approach using EMG and sound data to classify simulated bruxism behaviors under conscious conditions is proposed.

## Key findings

- Masseter EMG showed significantly higher discrimination accuracy in single-stream models.
- Multi-stream models did not show significant improvement over single-stream models.
- BMwoTC classification accuracy was below 0.5, indicating poor discrimination.

## Abstract

This study aimed to evaluate the potential use of machine learning to automatically classify electromyography (EMG) data into bruxism simulated movement with tooth contact (BMwTC), bruxism simulated movement without tooth contact (BMwoTC), and non‐bruxism movement (non‐BM).

Twelve eligible healthy participants (female/male: 2/10, mean age: 35.3 ± 8.4 years) were asked to perform the simulated movements (all the tasks were performed five times for 5 s each with a 30‐s rest interval). The electrodes were placed on the masseter, infrahyoid, inframandibular, and chin muscles. A sound sensor was placed adjacent to the masseter. The EMG and sound data were sampled at 1 and 44.1 kHz, respectively. Single‐ and multi‐stream hidden Markov models (HMMs) were used to automatically discriminate the tested behavior from the others using a hamming window with 100 ms and shift length of 50 ms. The leave‐one‐out method was used for training and testing the model, with data from 11 participants used for training and one for testing. Each participant was evaluated, and the final performance was measured by averaging the results of 12 classification trials. The validity of the discrimination was assessed by calculating the harmony mean values using six EMG signals and the sound data.

The masseter EMG demonstrated significantly higher discrimination accuracy in the single‐stream model (p  < 0.05, One‐way ANOVA, Tukey HDS). The multi‐stream model also demonstrated higher accuracy; however, no significant difference was observed. Notably, the accuracy of BMwoTC was less than 0.5.

The machine‐learning‐based discriminative system accurately discriminates BMwTC from non‐BM using masseter EMG.