Interaction- and asymmetry-aware facial blendshape analysis for objective quantification of Parkinsonian hypomimia

Abstract

Reduced facial expressivity (hypomimia) is an early motor feature of Parkinson’s disease (PD), yet its assessment still relies on subjective, coarse-grained clinical scales. Using ExpressionTracker, a consumer-grade application built on Apple’s ARKit framework, we derived 261 facial expression features during standardised cued expression tasks in 34 people with PD and 36 healthy controls (HC). Beyond the expected reduction in movement amplitude, PD was characterised by increased facial asymmetry and attenuated activity of ipsilateral mouth and eye regions. Interaction and asymmetry features were prominent among the between-group differences and explained 38 to 42 percent of the variance in clinician-rated hypomimia. For diagnostic classification (PD versus HC) under nested cross-validation, the best model (gradient boosting machine, GBM) reached an area under the receiver operating characteristic curve (AUC) of 0.834 (95% CI 0.733 to 0.923), followed by XGBoost (0.810, 95% CI 0.702 to 0.906) and light gradient boosting machine (LightGBM; 0.804, 95% CI 0.697 to 0.899). Engineered ARKit features outperformed an amplitude-only baseline (maximum AUC 0.75, 95% CI 0.63 to 0.86) and a demographic confounder-only baseline using age and sex (maximum AUC 0.68, 95% CI 0.55 to 0.80). These results indicate that automated facial blendshape analysis captures multidimensional motor dysfunction beyond amplitude reduction alone, positioning ExpressionTracker as a candidate digital biomarker for community screening, disease monitoring, and drug-efficacy assessment in future clinical trials.