An Optimized and Robust Machine Learning Framework for Early Parkinsons Disease Prediction Using Speech Signals

With the rapid advancement of technologies in the present era, predicting Parkinson's disease (PD) early using non-invasive and low-cost methods, such as speech analysis with machine learning (ML) tools, remains a challenging task and lacks sufficient confidence for healthcare providers to use in daily practice. Therefore, this study presents an optimized early PD prediction tool and investigates its stability and robustness using a rigorous evaluation mechanism. For early PD prediction using speech signal data, the eXtreme Gradient Boosting (XGB) model is optimized using the Tree-structured Parzen Estimator (TPE) method and the Synthetic Minority Oversampling Technique (SMOTE) for solving the imbalanced dataset problem. Its performance was rigorously evaluated using an optimized strategy to ensure reliability and to earn the trust of clinicians for real-world operational use. To validate the model's trustworthiness and prediction capability, it was evaluated through 10 different runs of Stratified 10-Fold Cross Validation (SCV). The average measures of accuracy as 96.76%, precision as 97.70%, f1-score as 96.70%, recall as 95.91% and ROC-AUC 98.72% show great progress and performance in comparison with similar works. The model performance and stability were evaluated in many different situations and showed that the proposed model is stable and strong enough, and could be used as a practical tool in daily medical care. This tool brings the opportunity to be used easily as a decision support system through a website and detect PD early using patient voice signal with low cost in a non-invasive way that could be used remotely and easily.