Efficient Voice-Based Parkinson Classification via Algorithm-Level Class Balancing.

Area of Science:

• Biomedical Engineering
• Machine Learning Applications
• Neurodegenerative Disease Research

Background:

• Parkinson's disease (PD) diagnosis relies on timely and accurate procedures for better patient outcomes.
• Machine learning shows promise for PD identification, but faces challenges with small, imbalanced voice datasets.
• Existing methods struggle with model generalization and stability due to data limitations.

Purpose of the Study:

• To objectively evaluate data preprocessing and model-level approaches for PD diagnosis under data constraints.
• To compare the effectiveness of different feature selection and class imbalance handling techniques.
• To identify optimal strategies for improving PD diagnosis using machine learning on voice data.

Main Methods:

• Three scenarios were developed: feature reduction with Recursive Feature Elimination and data resampling (Instance Hardness Threshold, K-means SMOTE, Borderline-SMOTE, SMOTE-Tomek).
• A third scenario employed Fisher score for feature selection and algorithm-level imbalance reduction using fine-tuned CatBoost and Support Vector Machine models.
• Models were trained and evaluated for PD classification using voice features.

Main Results:

• Data resampling methods (Scenario 1 & 2) reduced class imbalance but caused information loss and distorted decision boundaries.
• Fisher score feature selection combined with algorithm-level adaptation (Scenario 3) maintained data integrity.
• CatBoost in Scenario 3 achieved superior performance: 97% accuracy, 0.96 AUC, and 0.98 F1-score.

Conclusions:

• Feature-level refinement and algorithmic adaptation provide a robust approach for PD diagnosis.
• The CatBoost model, optimized with Fisher score and imbalance handling, demonstrates high efficacy.
• This study highlights the importance of tailored preprocessing and modeling for sensitive biomedical data.