Machine Learning-Based Comparison of 2-D and 3-D Approaches for Parkinson’s Disease Detection via Hand Movement Analysis

This study evaluates whether a conventional 2D smartphone webcam can serve as a viable and accessible alternative to a specialized 3D motion sensor (Leap Motion Controller) for the automated assessment of hypokinesia in Parkinson’s disease (PD). We conducted a cross-sectional study with synchronous recording of hand movements during three standardized MDS-UPDRS motor tasks—finger tapping (FT), hand opening/closing (OC), and forearm pronation/supination (PS)—in 32 PD patients and 10 healthy controls (HC). From the processed signals, we extracted 11 kinematic features quantifying amplitude, speed, frequency, rhythmicity, and decrement. We trained machine learning models (Logistic Regression, Random Forest, Support Vector Machines) to solve five clinical tasks: binary PD detection, early PD detection, Hoehn & Yahr staging (3- and 4-class), and MDS-UPDRS score classification. We performed a correlation analysis between matched 2D and 3D features and conducted a permutation feature importance analysis on a larger 3D dataset (119 PD patients and 27 HC). The 2D modality demonstrated strong, clinically relevant performance, particularly for binary classification tasks (66.4–75.5% balanced accuracy using Leave-One-Subject-Out validation). Key kinematic features, including average amplitude, speed, and frequency, showed high correlation (r ≥ 0.79 for FT/OC) between 2D and 3D modalities. These same features were identified as the most important predictors in the global feature importance analysis. While models trained on the larger 3D dataset achieved the highest accuracy (up to 82.4%), performance on the synchronously recorded 2D and 3D data was comparable, indicating the performance gap is attributable primarily to dataset size rather than a fundamental limitation of the 2D approach.