Quantitative Assessment of Facial Expression Asymmetry in Parkinson's Disease.

OBJECTIVE: Hypomimia, or reduced facial expressiveness, is a cardinal motor feature of Parkinson's disease (PD). Although limb motor symptoms in PD are characteristically asymmetric, whether facial hypomimia exhibits comparable asymmetry remains poorly understood. We investigated facial expression asymmetry in PD using AI-based computer vision and machine learning. METHODS: Videos of instructed emotional facial expressions were collected from 101 individuals with PD and 94 healthy controls (HCs) across two independent cohorts. Facial landmarks were extracted to compute region-specific facial asymmetry indices. Static and dynamic features derived from these indices were compared between groups and used to train machine learning models for PD-HC discrimination, with each model trained and evaluated separately within each cohort. Facial mobility was additionally quantified for each hemiface as the summed velocity of landmark displacement to assess concordance with limb motor laterality. RESULTS: Individuals with PD exhibited greater facial asymmetry than HCs, most prominently in the eyebrow and periocular regions during happy expressions (P = 0.01). Among participants with asymmetric limb parkinsonism, facial mobility was significantly reduced on the hemiface corresponding to the more affected limb (P = 0.002), indicating concordant facial-limb lateralization. Machine learning models trained on dynamic facial asymmetry features achieved up to 90.1% cross-validated accuracy in distinguishing PD from HCs. The models had comparable within-cohort performance on held-out test sets. CONCLUSIONS: Facial expression asymmetry is a measurable and lateralized feature of PD. Dynamic video-based analysis of facial asymmetry may offer an objective, scalable tool for assessing and monitoring facial motor function in PD.