Dynamic blinking feature extraction for automated facial nerve paralysis detection

Facial nerve paralysis (FNP) impair eyelid closure and blinking, risking ophthalmic complications and vision loss. Current detection methods primarily rely on static facial asymmetries, overlooking the dynamic eyelid movements during blinking that are important for evaluating treatment outcomes such as blink restoration. In this study, we present an automated system for objectively extracting dynamic blink features from high-frame-rate videos to address these limitations. We develop algorithms for dynamic blink feature extraction using a facial landmark detection model to capture eyelid movements and derive parameters for each blink. These parameters are processed with an Isolation Forest model to learn the typical distribution of combined parameters from both eyes, generating normality scores for each blink pair to indicate the degree of abnormality in upper eyelid movement while reducing noise from landmark detection and head movements. Our evaluation, which included 103 subjects (86 healthy and 17 with FNP), shows that the machine learning model trained to detect FNP using normality scores outperformed those trained with static parameters (with an increase of 75% in F1-score) and dynamic parameters (with an increase of 35% in F1-score). Notably, the normality score of the closing blink velocity, representing the speed at which the upper eyelid margin moves during the eye-closing phase, was the most distinguishing feature for FNP detection. These findings highlight the potential of the dynamic blink features in FNP detection and suggest further exploration to assess their effectiveness as objective measures for diagnosing FNP in addition to the facial asymmetry features proposed in other studies.