Sinhala language specific vocal biomarker extraction for parkinson’s diagnosis using machine learning

Abstract

Early detection of Parkinson's disease (PD) remains challenging, particularly in underrepresented populations where diagnostic resources are limited and culturally appropriate screening methods are needed. Limited research exists on using Sinhala voice biomarkers for PD detection in Sri Lankan populations through machine learning approaches. Speech data were collected from Sri Lankan PD patients and healthy controls, focusing on simple vowel sound articulation. Voice biomarker features were extracted and analyzed using four machine learning classifiers: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), and Logistic Regression (LR). Models were evaluated using precision, recall, accuracy, and F1-score metrics. A hybrid voting classifier combined the bestperforming algorithms. SVM and LR demonstrated superior performance among individual classifiers. The hybrid voting classifier combining these algorithms achieved 87% accuracy in detecting early-stage PD. Vocal biomarkers and hybrid machine learning strategies show promise for early PD recognition in marginalized populations.