Dynamic blinking feature extraction for automated facial nerve paralysis detection.

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Machine Learning

Background:

• Facial nerve paralysis (FNP) impairs eyelid function, increasing risks of ophthalmic complications and vision loss.
• Current FNP detection relies on static facial analysis, neglecting dynamic blink characteristics crucial for treatment assessment.

Purpose of the Study:

• To develop and validate an automated system for extracting dynamic blink features from high-frame-rate videos for objective FNP detection.
• To compare the efficacy of dynamic blink features against static parameters in diagnosing FNP.

Main Methods:

• Developed algorithms for dynamic blink feature extraction using facial landmark detection from high-frame-rate videos.
• Utilized an Isolation Forest model to generate normality scores for blink pairs, indicating upper eyelid movement abnormality.
• Evaluated the system on 103 subjects (86 healthy, 17 with FNP).

Main Results:

• The machine learning model using normality scores achieved a 75% higher F1-score than static parameters and a 35% higher F1-score than other dynamic parameters for FNP detection.
• The normality score of closing blink velocity emerged as the most significant feature for distinguishing FNP.
• The system demonstrated superior performance in objectively assessing blink dynamics.

Conclusions:

• Dynamic blink features offer a promising objective measure for FNP detection, complementing traditional facial asymmetry assessments.
• The developed automated system shows potential for improved diagnosis and monitoring of FNP.
• Further research is warranted to explore the full clinical utility of these dynamic blink metrics.