<![CDATA[Parkinson's disease (PD) is a progressive neurodegenerative disorder affecting more than 11.77 million people worldwide. Voice signal analysis has gained attention as a non-invasive screening approach because nearly 90% of PD patients experience measurable speech impairments. However, previous machine learning studies on PD voice datasets commonly face several limitations, including class imbalance that may lead to data leakage, the use of accuracy as the primary evaluation metric, and limited utilization of model interpretability methods. This study proposes a PD classification pipeline integrating SMOTE, XGBoost, and SHAP using the UCI Parkinson dataset, which consists of 195 samples and 22 acoustic features. A quantitative experimental approach was employed using 5-fold stratified cross-validation, where SMOTE was applied only to the training data within each fold to prevent data leakage, while SHAP was used for feature analysis and feature reduction experiments. The results showed that SMOTE improved the F1-Score from 0.9400 to 0.9527 and the Accuracy from 0.9077 to 0.9282. The final model achieved a mean AUC-ROC of 0.9614 and a Recall of 0.9592 across five folds. SHAP analysis showed differences between SHAP feature rankings and XGBoost built-in importance, with MDVP:Shimmer exhibiting the largest ranking change. In addition, the top-8 SHAP-ranked features achieved performance comparable to the full 22-feature model, obtaining an Accuracy of 0.9282 and an AUC of 0.9612. These findings indicate that the proper application of SMOTE and SHAP-based feature selection can improve model evaluation and provide additional information for feature analysis in Parkinson's disease classification.]]>
