Dynamic cheek surface modeling for enhanced hypomimia detection in Parkinson's disease.

Hypomimia, or facial masking, is a prominent symptom in Parkinson's disease (PD), impacting up to 90% of patients and reducing their quality of life by impairing facial expressivity. Detecting hypomimia is challenging due to its subtle onset and the limited availability of robust diagnostic tools for automatic analysis. To address these gaps, we propose a novel approach based on cheek surface variability, which captures subtle, progressive changes in cheek movement across different temporal granularities within video frames. To evaluate the effectiveness of our cheek marker approach, we also implemented and compared machine learning techniques based on geometric and surface features, and a deep learning approach using a 3D Convolutional Neural Network (3D CNN) to capture the temporal dynamics of facial movements. Additionally, this methodology includes a robust pre-processing pipeline to mitigate head rotation and posture biases, ensuring a consistent frontal face orientation throughout the analysis. Video samples of spontaneous speech were collected from 112 de novo, drug naive PD patients and 90 healthy control (HC) subjects. Experimental results demonstrate that the proposed cheek surface variability approach outperformed both the machine learning and deep learning models, achieving an unweighted averaged recall of 80.6% and emphasizing its effectiveness in distinguishing PD patients from HC subjects. This work highlights the importance of integrating temporal dynamics into hypomimia modeling and presents cheek variability as a valuable biomarker that enhances classification accuracy. It offers promising implications for more precise monitoring and diagnosis of PD.