CalibraPD-Ensemble: A Novel Non-Invasive Parkinson’s Disease Prediction Framework Using Calibrated and Explainable Hybrid Ensemble Learning from Biomedical Voice and Acoustic Biomarkers

Research overview

The text below is the source article’s original abstract under a verified reuse license. cc by

Parkinson’s disease is a progressive neurodegenerative disorder in which early and reliable detection remains essential for timely clinical management, patient monitoring, and therapy planning. Although molecular and genetic studies have improved understanding of Parkinson’s disease pathophysiology, there is also a growing need for non-invasive, interpretable, and clinically reliable predictive tools that can support diagnosis using accessible biomedical biomarkers. This study proposes a calibrated and explainable hybrid ensemble framework for Parkinson’s disease detection using two complementary biomedical voice and acoustic feature datasets. The proposed pipeline integrates dataset-specific preprocessing, nonlinear feature engineering, hybrid feature selection, conventional Machine Learning (ML), Deep Learning (DL), and three ensemble strategies, including meta-stacking, hybrid weighted fusion, and cascaded Decision Tree (DT). To enhance clinical reliability, the framework includes Bayesian hyperparameter optimization via Optuna, threshold optimization, statistical significance testing, bootstrap confidence estimation, probability calibration, and other eXplainable Artificial Intelligence (XAI) methods such as SHAP, LIME and permutation importance. The diagnostic performance is observed to be good in a heterogeneous dataset and is demonstrated in the experiment. The Hybrid Ensemble model had an AUC of 0.9924 and an AUPRC of 0.9976 for Dataset 1, and the highest accuracy and F1-score for the Cascaded DL model. The proposed framework showed competitive generalization performance with respect to different acoustic feature distributions on Dataset 2. The explainability analysis also identifies clinically relevant vocal and acoustic features indicative of Parkinson’s Disease (PD) prediction. Altogether, the results highlight the potential of non-invasive voice biomarkers combined with calibrated and interpretable AI models as useful decision-support evidence in screening, early prediction, and future personalized monitoring strategies for PD.

Related topics