Abstract
Parkinson’s disease (PD) is a progressive neurodegenerative disorder caused by the depletion of dopamine-producing neurons in the substantia nigra, leading to motor dysfunction and alterations in speech patterns. Temporal variations in speech signals have been identified as potential biomarkers of PD, emphasizing the importance of early detection for slowing disease progression and allowing timely therapeutic interventions. This study presents an optimized Convolutional Neural Network (CNN) classifier designed for accurate and efficient PD detection. The model leverages sustained vowel phonations (“ah”) from the Parkinson’s Disease Voice Recording Data Set, which includes samples from 147 patients with PD and 48 healthy controls. Sustained vowels are widely used in clinical protocols and research because they minimize articulatory variability and highlight relevant vocal biomarkers such as jitter, shimmer, and the harmonic-to-noise ratio (HNR). The preprocessing pipeline involved denoising, feature normalization, and Bayesian hyperparameter tuning to optimize performance. The CNN achieved a detection accuracy of 96.85%, along with high precision, sensitivity, and a well-balanced F1 score, demonstrating robust predictive capabilities. These findings validate the model’s reliability in detecting PD-specific voice patterns, even in noisy conditions, underscoring its potential for integration into clinical diagnostic workflows.
Funding Information
  • CAPES (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior), Brazil

This publication has 47 references indexed in Scilit: