Subject-Independent Parkinson’s Disease Detection Using Acoustic and Nonlinear Speech Features: A Robust Evaluation of Machine-Learning Classifiers

Abstract

Early detection of Parkinson’s disease can facilitate timely intervention. Speech is a noninvasive and cost-effective biomarker for Parkinson’s disease detection. In this study, a subject-independent speech-based pipeline was developed using acoustic and nonlinear speech features. To avoid subject-level information leakage, a strict fivefold subject-independent cross-validation procedure was performed. Eight machine-learning classifiers were evaluated using five speech feature representations: acoustic, spectral, nonlinear, hybrid, and standardized minimal acoustic features. Among these, the combination of standardized minimal acoustic features and logistic regression achieved the highest accuracy of 86.49%, with a sensitivity of 75.00% and a specificity of 95.24%. The hybrid feature representation achieved an accuracy of 83.78% and the highest area under the receiver operating characteristic curve (AUROC) of 0.8958. The nonlinear feature sets combined with tree-based and instance-based classifiers achieved accuracies of up to 81.08%. Error analysis showed that probabilistic classifiers produced more false-negative predictions, which is an important consideration for Parkinson’s disease screening. The findings indicate that the proposed pipeline provides a noninvasive and interpretable approach for Parkinson’s disease screening, with potential applications in remote speech-based assessment.