Table 1.
Summary of Parkinson’s disease (PD) classification studies.
Table 2.
Summary of Parkinson’s disease (PD) classification studies (continued).
Table 3.
Summary of Parkinson’s disease (PD) classification studies (continued).
Fig 1.
The overview of the Parkinson’s dataset analysis and explainable artificial intelligence.
Table 4.
Detailed description of the dataset attributes, including acoustic features and target labels.
Fig 2.
The overall preprocessing steps applied to the Parkinson’s dataset.
Fig 3.
Heatmap displaying the correlation between features in the dataset.
Fig 4.
Feature selection using Featurewiz: Selecting key features while removing redundant ones.
Fig 5.
Sampling distributions of class labels in the Base, SMOTE, and NearMiss versions of the dataset.
Table 5.
Performance metrics for base models: Class-wise precision, recall, F1 score, accuracy, G-mean, MCC, and STD.
Fig 6.
Regression analysis results for base models evaluated in the study.
Table 6.
Performance metrics for NearMiss models: Class-wise precision, recall, F1 score, accuracy, G-mean, MCC, and STD.
Fig 7.
Regression analysis results for models using Near Miss undersampling.
Table 7.
Performance metrics for SMOTE models: Class-wise precision, recall, F1 score, accuracy, G-mean, MCC, and STD.
Fig 8.
Regression analysis results for models using SMOTE oversampling.
Table 8.
Performance metrics for ensemble methods (RF, ADB, XGB) using base, SMOTE, and NearMiss.
Fig 9.
Visual comparison of the performance of the tuned KNN model with SMOTE, showcasing the confusion matrix and ROC curve.
(a) Confusion Matrix for Tuned KNN with SMOTE. (b) ROC Curve for Tuned KNN with SMOTE.
Table 9.
Comparison of the proposed model’s performance with existing research approaches.
Fig 10.
Receiver operating characteristic (ROC) curves for the nine models under (a) original imbalanced data, (b) NearMiss undersampling, and (c) SMOTE oversampling.
These curves illustrate each model’s classification performance across data balancing techniques. (a) Base Models. (b) Near Miss Models. (c) SMOTE Models.
Fig 11.
Precision-recall curves (PRC) for the nine models across different data balancing strategies: (a) original imbalanced dataset, (b) NearMiss undersampling, and (c) SMOTE oversampling.
These curves highlight precision-recall trade-offs under each condition. (a) Base Models. (b) Near Miss Models. (c) SMOTE Models.
Fig 12.
Comparison of original and predicted data demonstrating model effectiveness in capturing patterns.
Fig 13.
SHAP summary scatter plot showing the effect of individual feature values on the model’s output.
Red indicates high feature values, blue indicates low.
Fig 14.
SHAP summary bar plot showing average feature importance.
PPE, Fo(Hz), D2, and spread2 were the most influential in the best model.
Fig 15.
SHAP waterfall plot showing feature-level contributions for a specific prediction.
Features such as DFA, RPDE, and Fo(Hz) reduce the prediction score, while PPE slightly increases it.
Fig 16.
LIME explanation for an individual instance.
Features like spread2, PPE, and D2 contribute significantly to the model’s prediction, corroborating the SHAP results.