Benchmarking Video Foundation Models for Remote Parkinson’s Disease Screening

We present a large-scale, systematic benchmark of video foundation models (VFMs) for remote PD screening, built on a novel dataset of $\mathbf{1,888}$ participants ($727$ with PD) comprising $\mathbf{32,847}$ videos across $16$ standardized clinical tasks. Evaluating seven state-of-the-art VFMs — including VideoPrism, V-JEPA, ViViT, and VideoMAE — we find VideoPrism excels at visual speech kinematics and facial expressivity, V-JEPA is superior for upper-limb motor tasks, and TimeSformer remains highly competitive for rhythmic tasks like finger tapping, with AUCs of $\mathbf{76.4\text{–}85.3\%}$ across tasks.

Recommended citation: Islam, M.S., Hossain, E., Abdelkader, A., Adnan, T., Mashrur, F.R., Park, S. et al., 2026. Benchmarking Video Foundation Models for Remote Parkinson's Disease Screening. International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI).
Download Paper