Enhancing Parkinson's Disease Detection by Combining SMOTE and Feature Selection for Improved Machine Learning...

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Parkinson's disease (PD) diagnosis is challenging, especially early on.
• Voice analysis is a promising noninvasive method for PD detection.
• Class imbalance and high-dimensional data hinder machine learning (ML) accuracy.

Purpose of the Study:

• To optimize machine learning models for early Parkinson's disease detection using voice analysis.
• To evaluate the combined impact of data balancing and feature selection techniques.
• To identify the most effective preprocessing and classification methods for PD voice detection.

Main Methods:

• Applied Synthetic Minority Over-sampling Technique (SMOTE) for data balancing.
• Utilized feature selection (FS) methods: Analysis of Variance, Chi-squared (χ2), and Mutual Information.
• Integrated preprocessing with classifiers: XGBoost, Random Forest, Logistic Regression, and Support Vector Machine.

Main Results:

• Combining SMOTE with FS significantly improved ML model performance over individual techniques.
• XGBoost with Chi-squared (χ2) FS achieved the highest accuracy (96.4%) and F1-score (96.9%).
• Optimal performance was reached using approximately 600 selected voice features.

Conclusions:

• Effective preprocessing, including data balancing and feature selection, is crucial for accurate ML-based PD detection from voice.
• This approach supports the development of advanced, noninvasive diagnostic tools for Parkinson's disease.
• Optimized ML models show high potential for early and reliable PD identification.