AI analyzes facial motion to detect Parkinson's with 82.6% accuracy

Researchers at Majmaah University, led by Riyadh Almushrafy, have published a new paper on arXiv demonstrating that temporal facial-region motion analysis can effectively classify Parkinson's disease (PD) from in-the-wild YouTube videos. The study focuses on hypomimia (reduced facial expressivity), a common motor symptom of PD. By extracting geometric and velocity-based descriptors from 14 facial regions, the team compared multiple feature sets—static geometry, normalized geometry, velocity, relative velocity, and a GRU sequence baseline—under a binary classification protocol. The best performer used normalized velocity descriptors with a Random Forest classifier, achieving a balanced accuracy of 0.826 and an AUROC of 0.855 on the held-out test split. Over 10 random seeds, the representation remained stable (accuracy 0.810 ± 0.018, AUROC 0.855 ± 0.005), highlighting the robustness of this lightweight approach.

The work emphasizes interpretability through region-level ablation and permutation importance analysis, allowing researchers to see which facial regions contribute most to PD classification. Unlike complex deep learning models, this method is computationally efficient and transparent, making it suitable for remote screening applications. The authors frame the study as a benchmark-level analysis on the YouTubePD dataset, specifically noting it does not aim to assess clinical severity or replace MDS-UPDRS scoring. This research opens the door for scalable, non-invasive Parkinson's detection tools using standard webcams, potentially enabling earlier diagnosis and monitoring in underserved populations. The paper has been submitted to Biomedical Signal Processing and Control.