Predicting facial deformation during respiratory mask fitting with semi-supervised graph neural networks

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Computational Mechanics

Background:

• Prolonged respiratory mask use causes discomfort and pressure sores due to poor fit.
• Inadequate mask fit compromises protection effectiveness and user compliance.
• Optimizing mask fit is crucial for user comfort, safety, and adherence to recommendations.

Purpose of the Study:

• To predict facial deformation and pressure distribution for optimal respiratory mask fit.
• To develop a personalized respiratory mask model using limited biomechanical data.
• To enhance comfort, safety, and compliance in medical and industrial mask applications.

Main Methods:

• Designed a semi-supervised graph neural network representing facial geometries as graph structures.
• Utilized a variational graph autoencoder constrained by Hertzian contact theory.
• Incorporated an XGBoost module for deformation zone classification with 45 labeled and 120 unlabeled facial datasets.

Main Results:

• Achieved 0.164 mm deformation RMSE (R2=0.9896) and 0.0492 kPa pressure RMSE (R2=0.9517).
• Demonstrated significant improvements over baseline models: 34.27% over Random Forest, 20.62% over PointNet++, and 10.01% over TPSNET in R2.
• Confirmed robust generalization with minimal overfitting via five-fold cross-validation and achieved sub-2-second inference time.

Conclusions:

• Introduced a real-time personalized respiratory mask model for precise prediction of facial deformation and contact pressure.
• The approach effectively generalizes from limited labeled data, improving mask comfort and safety.
• The developed model enhances compliance with mask-wearing recommendations in diverse environments.