Fig. 1: Overview of workflows for system-building, validation, and video analysis.

For the system-building stage, each clip was scored independently by a movement disorder specialist. At the same time, the hand skeleton in each clip was extracted using a combination of MediaPipe and preprocessing methods (including normalization and null value processing) to ensure the quality of the hand keypoints for model training and testing. Then, the hand keypoints can be used in deep learning model building and quantitative hand parameters calculation. The hyperparameters were optimized by grid-searching and data augmentation (3D keypoint rotation and random cropping). A well-trained model was picked to estimate the MDS-UPDRS item score. Furthermore, four hand parameters were calculated and compared with the estimated MDS-UPDRS item score to verify and interpret the model. For the video-analyzing stage, an inferencing pipeline was built, including the keypoint transformation, quantitative hand parameters calculation and MDS-UPDRS item score estimation. The model was also verified by an outer validation dataset. Lastly, users can access the service to assess their motor movement via the website interface.