Vocal Biomarkers for Parkinson's Disease Classification Using Audio Spectrogram Transformers.

Area of Science:

• Neurology
• Artificial Intelligence
• Speech Science

Background:

• Parkinson's disease (PD) is a neurodegenerative disorder impacting motor and non-motor functions, including speech.
• Vocal biomarkers offer a potential non-invasive method for PD detection.
• Traditional deep learning models have limitations in capturing subtle speech impairments in PD.

Purpose of the Study:

• To evaluate the effectiveness of the Audio Spectrogram Transformer (AST) model for detecting Parkinson's disease using vocal biomarkers.
• To compare AST's performance against established deep learning architectures for PD detection.
• To assess the AST model's capability for cross-lingual generalization in PD voice analysis.

Main Methods:

• Speech recordings from 150 participants (PD and healthy controls) across two datasets (PC-GITA and ITA) were analyzed.
• The Audio Spectrogram Transformer (AST) model was employed and compared with VGG16, VGG19, ResNet18, ResNet34, vision transformer, and swin transformer.
• Standard audio preprocessing included sampling rate standardization to 16 kHz and amplitude normalization.

Main Results:

• The AST model achieved superior classification accuracy: 97.14% (ITA), 91.67% (PC-GITA), and 92.73% (combined dataset).
• AST outperformed traditional architectures by 5%-10% in accuracy, demonstrating robust cross-lingual generalization.
• Consistent performance was observed across speech tasks, with high precision (0.97) and recall (0.96) in sustained vowel analysis.

Conclusions:

• The AST model provides a reliable and non-invasive method for Parkinson's disease detection through voice analysis.
• The model's strong cross-lingual generalization suggests potential for broad clinical application.
• Further validation across diverse populations is recommended for clinical implementation of AST-based PD detection.