A Community Engaged Open Science Model for Voice-Based Parkinson's Disease Detection Using Advanced Machine Learning Techniques

The task of classifying Parkinson’s disease (PD) using vocal signal features is hindered by issues such as high- dimensional input space and imbalanced class distribution, both of which can limit the robustness and generalizability of machine learning models. To overcome these barriers, this study presents an advanced diagnostic framework that combines dimensionality reduction through Principal Component Analysis (PCA), tar- geted data resampling via the Synthetic Minority Oversampling Technique (SMOTE), and a cascade-based ensemble learning architecture. The proposed methodology incorporates stratified train-test splitting, Bayesian hyperparameter optimization using Optuna, and a two-level stacking structure featuring Logistic Regression and Gradient Boosting as meta-learners. Fifteen base classifiers and multiple ensemble strategies including bagging, voting, stacking, a deep neural ensemble with static weight averaging, and a novel cascaded design are systematically eval- uated on a dataset of 195 voice samples. The cascaded ensemble achieves a peak performance of 97.44% accuracy and an F1- score of 98.25%, surpassing traditional models in both predictive strength and reliability. Noteworthy contributions include the exclusive application of SMOTE to training data after the split to mitigate data leakage, PCA-based feature compression while retaining 95% of the variance, and the efficient tuning of model parameters through Optuna’s Bayesian search algorithm. These results underscore the value of strategically integrated pre- processing and ensemble techniques in enhancing classification accuracy and minimizing overfitting, suggesting a promising path forward for reliable, and accessible voice-based screening tools for Parkinson’s disease.