Figure 1.

Overview of the proposed multimedia data quality assurance framework. (1) Three experts first independently rated audio/video recordings of finger-tapping, facial expression, and speech tasks. Through collaborative discussion on 100+ samples per task, they established task-specific annotation guidelines and created a consensus-labeled test set. (2) Using these guidelines, experts independently annotated additional recordings to build the training dataset. (3) We then engineered explainable features aligned with the guidelines and trained machine learning models to predict data quality. High-quality data can be passed to the movement disorder detection model, while low-quality data may trigger Shapley additive explanations (SHAP)–based feedback highlighting key issues and prompting users to rerecord. A/V: audio/video; ML: machine learning.