Abstract: Background　The motor function assessment of Parkinson's disease has long relied on the subjective Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS) Part Ⅲ , which is susceptible to the rater's subjective experience. Objective　To develop a multidimensional machine learning model integrating general demographics, physical examination, and quantitative motor parameters for objectively predicting MDS-UPDRS scores in finger-tapping tasks of patients with Parkinson's disease. Methods　This was a single‑center, model development and validation study. Patients with Parkinson's disease consecutively admitted to the Department of Neurology, the First Medical Center of PLA General Hospital from January 10 to December 10, 2024 were enrolled. Demographic data were collected, physical parameters such as the distance from the fifth metacarpal base to the distal wrist crease were measured, and a computer‑vision‑based motion assessment system was used to record videos of 10 consecutive index‑to‑thumb taps, from which kinematic parameters including frequency, amplitude, and their coefficients of variation were extracted. The primary outcome was the MDS‑UPDRS finger tapping‑tapping item score (ordinal multiclass, 0 ‑ 4) determined by three certified neurologists through blinded video review. The least absolute shrinkage and selection operator (LASSO) regression combined with group cross validation was used to select the key features. Seven mainstream algorithms were compared: k‑nearest neighbors (KNN), decision tree (DT), ordinal logistic regression (OLR), random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). Model performance was evaluated using leave‑one‑out cross validation (LOOCV). Performance was assessed using the multiclass Kappa coefficient, accuracy, area under the curve (AUC), and confusion matrix. Feature importance was analyzed using the Shapley Additive exPlanations (SHAP). Results　A total of 81 patients were recruited, 5 were excluded, and finally 76 patients (152 tested hands) were analyzed. LASSO regression retained the following predictors: frequency, average amplitude, coefficient of variation of duration, distance from the fifth metacarpal base to the distal wrist crease, thumb length, sex, coefficient of amplitude variation, and standard deviation of average amplitude difference. LOOCV showed that the ordinal logistic regression model performed best in the multiclass task (0 ‑ 4): AUC was 0.893 (95% CI: 0.859 ‑ 0.928), Kappa coefficient was 0.542 (95% CI: 0.410 ‑ 0.665), and accuracy was 66.5% (95% CI: 57.2% ‑ 75.0%). SHAP feature importance analysis revealed that frequency (37.3%), average amplitude (20.5%), and the fifth metacarpal base to the distal wrist crease (9.1%) were the three most contributing features. Conclusion　This study successfully develops and preliminarily validates a multimodal machine learning model, confirming that integrating individual ,static anthropometric features can improve the objective prediction of finger-tapping motor function scores. This model provides a new methodological pathway for automated and quantitative motor assessment in Parkinson's disease.