Deep learning and predictive modelling for generating normalised muscle function parameters from signal images of...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Computational Neuroscience

Background:

• Proprietary software limits access to archived mandibular electromyography (EMG) signals.
• A lack of open-source tools hinders analysis of EMG data for temporomandibular joint (TMJ) complex function.
• Pattern recognition and predictive modeling for TMJ disorders require accessible, normalized EMG data.

Purpose of the Study:

• To develop an open-source workflow for extracting normalized signal parameters from mandibular EMG images.
• To identify optimal clustering methods for quantifying EMG signal intensity and activity durations.
• To enable continuous access and analysis of mandibular EMG data for research.

Main Methods:

• A workflow using OpenCV, variational encoders, and Neurokit2 generated and augmented 866 unique EMG signals from jaw movement exercises.
• K-means, Gaussian Mixture Model (GMM), and DBSCAN clustering algorithms were employed for signal normalization and processing.
• The workflow was validated using EMG data from 66 participants, measuring temporalis, masseter, and digastric muscles.

Main Results:

• The developed workflow successfully extracted normalized signal data from EMG images.
• Quantifiable parameters for muscle activity duration and functional intensity were generated.
• K-means clustering performed best for temporalis and masseter muscles during chewing (silhouette scores 0.10-0.11), while DBSCAN and GMM showed lower scores for other movements.

Conclusions:

• A novel deep learning workflow enables the extraction and quantification of mandibular EMG data from images.
• This approach overcomes limitations of proprietary software, facilitating advanced analysis of TMJ complex function.
• The study provides a valuable tool for researchers studying jaw muscle activity and related disorders.