Video-Based Quantitative Assessment of Facial Expressivity in the Follow-up of Parkinson’s Disease
Video-Based Quantitative Assessment of Facial Expressivity in the Follow-up of Parkinson’s Disease
Publication status: preprint
A plain-language reading has not been prepared for this paper yet.
Original abstract
Facial expressivity analysis in Parkinson’s disease (PD) has received considerable attention due to the characteristic symptom of hypomimia, which is defined as the reduction or loss of spontaneous facial movements and emotional facial expressions. Hypomimia has been linked to a detrimental impact on close relationships and psychological well-being. This investigation presents the preliminary findings of a quantitative analysis of facial expressivity in a longitudinal pilot cohort of nine Mexican subjects diagnosed with PD, evaluated during two sessions: at baseline and following eight weeks of clinical follow-up under stable medication conditions. In each session, participants were instructed to mimic eight specific facial expressions while their faces were recorded with a high-definition camera. Facial expressivity was quantified by extracting dynamic features (Euclidean distances) between key landmarks from video recordings using the MediaPipe FaceMesh framework (utilizing 468 3D landmarks, excluding the iris). The landmark coordinates were normalized for translation, scale (via intercanthal distance), and rotation. Our quantitative results showed that the normalized landmark distances had a higher offset (representing the resting baseline) and a wider peak-to-peak amplitude at the baseline evaluation compared to the eight-week follow-up, indicating a significant reduction in facial mobility and expressivity. These physiological findings concurred with a decrease in the participants’ clinical status, as assessed by the Unified Parkinson’s Disease Rating Scale (UPDRS) Activities of Daily Living (ADL) and Motor Examination Scores (MES). To demonstrate the discriminative power of the extracted dynamic features, a Support Vector Machine (SVM) classifier was evaluated using a Leave-One-Subject-Out (LOSO) cross-validation scheme to prevent data leakage. The SVM model successfully distinguished between baseline and follow-up sessions with an exploratory participant-independent accuracy of 71.4% (F1-score of 70.6%). This pilot study demonstrates the feasibility of video-based, non-invasive digital tools to track subtle longitudinal changes in hypomimia, highlighting their potential as complementary, objective instruments for disease monitoring and clinical follow-up.