A hybrid computational framework for Parkinson’s disease prediction using fractional-order modeling and machine learning via vocal biomarkers

This study introduces a hybrid approach combining fractional-order dynamical modeling with machine learning (ML) to predict Parkinson’s disease (PD) using vocal biomarkers. A fractional-order system captures memory-dependent vocal impairment dynamics, with proven stability and uniqueness. Calibrated with clinical data, it achieves 97.44 % accuracy, while a Deep Neural Network (DNN) attains 91.8 % accuracy (AUC: 95.8 %). Key biomarker Pitch Period Entropy (PPE) (importance: 0.9014) emerges as most discriminative. The fusion of fractional calculus and ML enhances interpretability and predictive power, enabling early, non-invasive PD detection. This framework advances computational neurology, offering potential for personalized monitoring and improved diagnostic tools.