A Comparative Review of Machine Learning Models for Parkinson’s Disease Classification

Abstract

Parkinson’s disease (PD) is a chronic, progressive neurological condition that disrupts motor and non-motor abilities, underscoring the urgent need for early and accurate detection to enable prompt therapeutic measures. Although many studies report near-perfect accuracy on the widely used Oxford Parkinson’s voice dataset, these figures usually come from a single small hold-out split, without uncertainty quantification, statistical testing or interpretability, which makes them hard to compare and to translate into clinical practice. This study therefore proposes a rigorous and reproducible benchmarking protocol for logistic regression, decision trees, k-nearest neighbors, support vector machines, multilayer perceptrons, Random Forests, XGBoost, LightGBM, CatBoost, as well as voting and stacking ensembles. The protocol combines a subject-disjoint 75/25 hold-out split, 5-fold subject-grouped cross-validation with SMOTE re-fitted inside each fold, training-only preprocessing and augmentation, subject-level bootstrap confidence intervals, McNemar tests with Holm correction (including a subject-clustered variant), clinical error analysis, computational cost and SHAP-based explainability.The stacking ensemble achieved the best and most stable cross-validated performance (accuracy 93.15 ± 3.78%, AUC 98.00 ± 0.76%, sensitivity 96.36 ± 3.40%) and, with Random Forest, CatBoost, XGBoost and soft voting, reached 95.92\% test accuracy (95% CI: 91.67--100%), 97.30% sensitivity and 91.67% specificity, i.e., one missed PD recording out of 37. None of the 55 pairwise differences remained significant after Holm correction, showing that the top models form a statistically indistinguishable tier on this small cohort, in which fundamental frequency (MDVP:Fo), pitch period entropy (PPE), shimmer (APQ5) and detrended fluctuation analysis (DFA) were the most influential vocal biomarkers.This study examines the strengths, weaknesses, and diagnostic value of modern machine learning methods while considering key challenges such as class imbalance, generalization, data leakage, and interpretability. It provides useful guidance for selecting appropriate ML tools for Parkinson’s disease screening and helps pave the way for future progress in multimodal systems and explainable artificial intelligence for healthcare applications.