Vision Transformer for Parkinson's Disease Classification using Multilingual Sustained Vowel Recordings

Area of Science:

• Neuroscience
• Computational Linguistics
• Biomedical Engineering

Background:

• Parkinson's disease (PD) is the second most common neurodegenerative disorder globally.
• Early detection of PD is critical for timely intervention and management.
• Voice and speech disorders affect a significant majority (70-90%) of PD patients.

Purpose of the Study:

• To develop and validate a novel pipeline for automatic Parkinson's disease classification.
• To explore the potential of sustained vowel recordings and mel-spectrograms for PD detection.
• To assess the efficacy of a Vision Transformer model for language-independent voice analysis in PD.

Main Methods:

• Utilized a Vision Transformer model applied to mel-spectrograms derived from multilingual sustained vowel recordings.
• Developed a classification pipeline for distinguishing individuals with and without Parkinson's disease.
• Evaluated the model's performance using the F1-score metric.

Main Results:

• The proposed Vision Transformer model achieved an F1-score of 0.78 for Parkinson's disease classification.
• The model demonstrated effectiveness as a single-modality biomarker for PD detection, irrespective of language.
• The results align with the high prevalence of voice disorders in Parkinson's disease patients.

Conclusions:

• Voice analysis using advanced deep learning models like Vision Transformers shows significant potential for early and accurate Parkinson's disease diagnosis.
• The developed method offers a non-invasive, language-independent approach suitable for widespread clinical application and telemedicine.
• This approach can facilitate faster diagnosis, guide treatment initiation, and aid in risk prediction for Parkinson's disease.