A Unified Deep Learning Ensemble Framework for Voice-Based Parkinson's Disease Detection and Motor Severity...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Computational Neuroscience

Background:

• Parkinson's disease (PD) diagnosis relies on clinical assessment, often subjective.
• Biomedical voice analysis offers a non-invasive method for PD detection and monitoring.
• Accurate prediction of motor severity is crucial for personalized PD management.

Purpose of the Study:

• To develop a hybrid ensemble learning framework for simultaneous PD detection and motor severity prediction.
• To leverage deep multimodal fusion and advanced feature engineering for enhanced performance.
• To establish a scalable foundation for voice-based clinical decision support systems for PD.

Main Methods:

• A hybrid ensemble learning framework integrating deep multimodal fusion with expert pathways and self-attention.
• Preprocessing included outlier removal, two-stage feature scaling, and data augmentation.
• Ensemble strategies involved stacking fusion model outputs with tree-based regressors (Random Forest, Gradient Boosting, XGBoost).

Main Results:

• The Stacking Ensemble with XGBoost (SE-XGB) achieved 99.78% R² and 0.3802 RMSE for UPDRS regression.
• The framework attained 99.37% accuracy for Parkinson's disease classification.
• Comparative analysis demonstrated superior performance over recent literature, especially in regression.

Conclusions:

• The proposed hybrid ensemble framework effectively integrates deep learning and ensemble meta-modeling for voice-based PD monitoring.
• This approach provides accurate and generalizable models for both PD detection and motor severity prediction.
• The study offers a scalable solution for future clinical decision support systems in Parkinson's disease management.