Bradykinesia, a core feature of Parkinson’s disease (PD), often emerges early in disease progression but remains challenging to quantify objectively. Conventional assessments rely on visual scoring by experts, which is subjective, time-consuming, and difficult to scale. We developed and validated an AI-driven, video-based system for automated detection of PD from short recordings of the finger-tapping task. Videos from 51 people with PD (pwPD, rated as normal or having slight motor dysfunction by a trained clinician on the MDS-UPDRS finger-tapping item) and 43 healthy controls were collected across 15 clinical sites under non-standardized conditions and analyzed using the open-source VisionMD software. Normalized kinematic time-series and multiple bradykinesia-related features were extracted. We trained and compared interpretable feature-based classifiers and time-series-based classifiers using nested cross-validation, bootstrap analysis, and decision-curve evaluation. The feature-based Gradient Boosting model achieved the best performance (ROC-AUC = 0.94 ± 0.03), outperforming the MultiRocket time-series model (ROC-AUC = 0.85 ± 0.05). Feature selection identified seven physiologically meaningful predictors related to movement speed, decay, and variability. Group-level analyses confirmed significant reductions in amplitude and velocity and increased variability among pwPD, consistent with early bradykinesia and the sequence effect. AI-based video analysis can accurately detect PD-related motor alterations even in individuals with minimal clinical signs. By quantifying subtle velocity and rhythmicity deficits from brief, smartphone-quality videos, this approach enables objective, scalable early screening, supporting equitable access to specialist-level evaluation and precision disease management.
Guarín et al. (Mon,) studied this question.