Early Parkinson’s Detection via Aggregated Acoustic Features and Ensemble Learning

Abstract

Parkinson’s disease (PD) is progressive neurological disorder. The early detection is critical for symptoms reduction. Though early diagnosis is a challenging task even for trained doctors, as studies reported around 25% of Parkinson patients were mistakenly diagnosed in their early stages. So, this study proposes an ensemble machine learning model that trained on 240 voice samples. The 240 audio samples were a mix between sick and healthy 80 persons. The subject-level splitting with Group K-Fold across validation was the first approach, while the other was an advanced feature aggregation. The performance of the models with the stacked of ensemble of Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP) is greatly improved by the aggregation which obtained the highest level of accuracy (88.75%) and F1-score (88.88%). This aggregation method increased robustness by lowering intra-subject variability, while ensemble learning used complementary model to improve classification. The results showed that the meta classifier’s complementary capabilities enable it to exploit various decision boundaries, resulting in enhanced generalization and higher diagnostic performance in detecting PD.