A Split-Face Computer Vision/Machine Learning Assessment of Facial Paralysis Using Facial Action Units

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Machine Learning

Background:

• Subjectivity and variability in clinician-graded scales limit accurate facial paralysis assessment.
• Objective and reproducible methods are needed for patient management and research in facial paralysis.

Purpose of the Study:

• To evaluate an automated video analysis pipeline for classifying facial paralysis severity.
• To validate the pipeline against expert clinical grading using hemifacial mirroring and Facial Action Unit (AU) quantification.

Main Methods:

• A custom Python pipeline processed 112 clinical videos, generating mirrored hemifacial constructs.
• OpenFace 2.0 extracted AU intensities, and XGBoost models predicted paralysis severity (None, Incomplete, Complete) per facial zone.
• Models were trained and validated against multi-expert clinical assessments.

Main Results:

• The automated pipeline achieved high accuracy in predicting facial paralysis severity across facial zones (upper face: 0.83, mid-face: 0.93, lower face: 0.84).
• Mirrored AU intensities showed significant differences across expert-defined severity groups, confirming feature relevance.
• Weighted F1-scores ranged from 0.82 to 0.92, indicating robust performance.

Conclusions:

• An automated pipeline utilizing hemifacial mirroring and mirrored AU analysis accurately predicts facial paralysis severity from standard clinical videos.
• This technology offers a promising objective tool for facial paralysis assessment, potentially improving clinical decision-making and research outcomes.