Convolutional neural network ensemble for Parkinson's disease detection from voice recordings

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Speech Pathology

Background:

• Parkinson's disease (PD) diagnosis can be improved with objective measures.
• Dysarthria, a speech disorder, is an early PD symptom.
• Deep learning models offer automated feature extraction for voice analysis.

Purpose of the Study:

• To develop and evaluate an ensemble of convolutional neural networks (CNNs) for Parkinson's disease detection using voice recordings.
• To assess the efficacy of a multiple-fine-tuning method for training CNNs on voice data.
• To determine the performance of vowel-specific voice analysis for PD detection.

Main Methods:

• Utilized voice recordings from 50 healthy individuals and 50 PD patients from the PC-GITA database.
• Employed an ensemble of CNNs trained with a novel multiple-fine-tuning approach.
• Performed 10-fold cross-validation, analyzing each vowel (/a/, /e/, /i/, /o/, /u/) separately.

Main Results:

• The CNN ensemble successfully differentiated between PD patients and healthy individuals across all tested vowels.
• The best performance was achieved using the /a/ vowel, yielding 99% accuracy, 86.2% sensitivity, 93.3% specificity, and 89.6% AUC.
• The model demonstrated robust performance, indicating its potential for clinical application.

Conclusions:

• The proposed deep learning method shows significant potential for the early screening, diagnosis, and monitoring of Parkinson's disease via voice analysis.
• Vowel-based voice recordings are a feasible and accessible method for PD detection, requiring no specialized hardware.
• This approach offers a non-invasive, objective tool to complement existing clinical practices for Parkinson's disease management.