A Robust Voice-Based Parkinson's Disease Diagnosis Model via Semantic RGB Feature Transformation and Hybrid Deep...

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Early diagnosis of Parkinson's disease (PD) is crucial for effective patient management and timely intervention.
• Current diagnostic methods can be invasive or lack sensitivity in early stages.
• Voice analysis offers a non-invasive approach for PD detection.

Purpose of the Study:

• To develop and validate a non-invasive, AI-supported hybrid diagnostic model for early Parkinson's disease detection using voice recordings.
• To transform acoustic features into a structured RGB image format for deep visual representation learning.
• To evaluate the performance of the proposed model compared to existing machine learning and deep learning approaches.

Main Methods:

• A hybrid AI model was developed using acoustic features from voice recordings of 188 PD patients and 64 healthy individuals.
• TQWT-based, MFCC/delta/log-energy, and other acoustic features were semantically grouped and mapped to RGB color channels.
• The generated RGB images were processed using the Pyramid Vision Transformer (PVT) model, with extracted features further classified by SVM, random forests, logistic regression, and majority vote classifiers.
• Subject-level grouping was implemented to prevent data leakage during evaluation.

Main Results:

• The proposed hybrid model achieved high diagnostic performance: 96.18% accuracy, 98.24% sensitivity, 90.16% specificity, 96.68% precision, and 97.44% F1 score.
• Ablation studies demonstrated superior performance compared to direct ViT_Base, Swin Transformer classification, and traditional machine learning methods using raw features.
• Performance gains were primarily attributed to the semantic RGB representation and PVT-based deep feature extraction.

Conclusions:

• The developed AI framework provides a potentially applicable and scalable solution for voice-based clinical decision support in Parkinson's disease screening.
• The novel approach of transforming acoustic data into RGB images for deep learning shows significant promise for early PD detection.
• This non-invasive method could enhance early diagnosis, leading to improved patient outcomes and management strategies.