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

This study presents a hybrid ensemble learning framework for the joint detection and motor severity prediction of Parkinson's disease (PD) using biomedical voice features. The proposed architecture integrates a deep multimodal fusion model with dense expert pathways, multi-head self-attention, and multitask output branches to simultaneously perform binary classification and regression. To ensure data quality and improve model generalization, preprocessing steps included outlier removal via Isolation Forest, two-stage feature scaling (RobustScaler followed by MinMaxScaler), and augmentation through polynomial and interaction terms. Borderline-SMOTE was employed to address class imbalance in the classification task. To enhance prediction performance, ensemble learning strategies were applied by stacking outputs from the fusion model with tree-based regressors (Random Forest, Gradient Boosting, and XGBoost), using diverse meta-learners including XGBoost, Ridge Regression, and a deep neural network. Among these, the Stacking Ensemble with XGBoost (SE-XGB) achieved the best results, with an R² of 99.78% and RMSE of 0.3802 for UPDRS regression and 99.37% accuracy for PD classification. Comparative analysis with recent literature highlights the superior performance of our framework, particularly in regression settings. These findings demonstrate the effectiveness of combining advanced feature engineering, deep learning, and ensemble meta-modeling for building accurate and generalizable models in voice-based PD monitoring. This work provides a scalable foundation for future clinical decision support systems.