Spectro-Image Analysis with Vision Graph Neural Networks and Contrastive Learning for Parkinson's Disease Detection.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Speech Processing

Background:

• Parkinson's disease (PD) diagnosis relies on clinical symptoms, often diagnosed late.
• Speech analysis offers a non-invasive method for early PD detection.
• Current methods struggle to capture complex spectro-temporal speech alterations in PD.

Purpose of the Study:

• To develop an advanced framework for Parkinson's disease detection using speech.
• To enhance spectro-temporal speech analysis through novel machine learning integration.
• To improve the accuracy and viability of speech-based PD diagnostic tools.

Main Methods:

• A novel framework integrating Vision Graph Neural Networks (ViGs) with supervised contrastive learning.
• Frequency band decomposition of audio signals into low, mid, and high-frequency spectral representations.
• Processing mel multi-band spectro-temporal representations using a ViG architecture trained with a contrastive objective.

Main Results:

• The ViG-contrastive framework demonstrated superior classification performance in PD detection.
• The ViG-M-GELU architecture achieved a test accuracy of 91.78%.
• The framework effectively learned discriminative representations from limited labeled data.

Conclusions:

• The proposed ViG-contrastive framework offers a promising approach for accurate Parkinson's disease detection.
• This method captures complex spectro-temporal speech relationships missed by traditional Convolutional Neural Network (CNN) approaches.
• The study highlights the potential for developing clinically viable, speech-based diagnostic tools for PD.