Impact of Spectral Representation and Vowel Selection on Graph Neural Network Performance for Parkinson's Disease Detection from Speech

Abstract

Parkinson's illness (PD) is a progressive neurodegenerative condition that often manifests as hypokinetic dysarthria before severe motor disability. The early non-invasive detection has clinical significance from sustained vowel speech analysis; However, the existing deep learning based approaches use convolutional or recurrent architectures that do not explicitly capture the inter-segment relational structure. The goal of this paper is to present the design of a Graph Neural Network (GNN) approach for the detection of PD from sustained vowels. It systematically studies the effect of the spectral representation (Mel vs Log-Mel spectrograms) as well as the phonetic content (five sustained vowels: /a/, /e/, /i/, /o/, /u/) on classification performance. The spectrogram segments are treated as graph nodes, with edges weighted by cosine similarity and pruned. A binary impulse versus healthy classification is performed using a two-layer graph convolutional network with global mean pooling. The public clinical voice dataset is used on 195 recordings and 54 subjects using stratified five-fold cross validation to assess the experimental. Log-Mel representations significantly outperform a standard Mel spectrogram for all vowels (Wilcoxon p < 0.05).  Vowel /a/ showed the maximum discrimination, with AUC 0.863±0.021, accuracy 87.4%, sensitivity 76.1%, specificity 91.2%, and F1-Score 0.814 with Log-Mel features. The proposed framework is proven robust and interpretable through the study of graph threshold sensitivity, embedding separability and training convergence. The team shows that Log-Mel spectrograms plus graph-based relational modeling represent a reproducible, interpretable, and clinically relevant approach to PD speech detection.