Hybrid preprocessing and ensemble classification for enhanced detection of Parkinson's disease using multiple speech...

Area of Science:

• Biomedical Engineering
• Computational Linguistics
• Neurology

Background:

• Parkinson's disease (PD) prevalence is rising globally.
• Acoustic recording databases for PD are increasingly available.
• Machine learning (ML) offers potential for PD detection from speech.

Purpose of the Study:

• Evaluate the feasibility of ensemble ML for PD detection.
• Utilize diverse acoustic datasets for robust analysis.
• Identify key acoustic features indicative of PD.

Main Methods:

• Employed a hybrid preprocessing framework: RobustScaler (scaling), ROS/SMOTE/RUS (sampling), XGBoost/AdaBoost (classification).
• Tested models on three public PD speech datasets: MIU, UEX, and UCI.
• Evaluated performance using accuracy, precision, recall, and F1-score; utilized SHAP analysis for feature importance.

Main Results:

• Optimal preprocessing varied by dataset, but RobustScaler, combined sampling, and XGBoost/AdaBoost generally yielded superior results.
• Achieved high performance: MIU dataset (97.37% accuracy), UEX and UCI datasets (100% accuracy).
• SHAP analysis identified Mel-frequency cepstral coefficients as significant PD-related acoustic features.

Conclusions:

• Ensemble ML approaches are feasible for PD detection via acoustic recordings.
• Dataset-specific preprocessing strategies are crucial for optimal performance.
• Identified key acoustic features provide guidance for developing voice-based PD screening tools.