Classifying Parkinson's Disease Based on Acoustic Measures Using Artificial Neural Networks

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neurology

Background:

• Parkinson's disease (PD) diagnosis relies on clinical symptoms, often leading to delayed detection.
• Voice analysis offers a non-invasive method for potential early PD detection.
• Artificial neural networks (ANNs) are increasingly utilized for complex prediction tasks.

Purpose of the Study:

• To evaluate the efficacy of various feed-forward artificial neural networks (ANNs) for Parkinson's disease (PD) prediction.
• To assess the impact of different feature selection techniques on diagnostic accuracy.
• To identify the most informative voice features for PD detection.

Main Methods:

• Utilized multiple feed-forward artificial neural networks (ANNs) with diverse configurations.
• Extracted features from 26 distinct voice samples per individual.
• Applied feature selection methods including Pearson's and Kendall's correlation coefficients, principal component analysis (PCA), and self-organizing maps (SOMs).
• Validated results using the leave-one-subject-out (LOSO) cross-validation scheme.

Main Results:

• Multiple ANNs demonstrated high classification accuracy for PD diagnosis even without feature selection.
• Kendall's correlation coefficient-based feature selection yielded the best performance, identifying key voice features.
• A fine-tuned neural network achieved a test accuracy of 86.47% for PD prediction.

Conclusions:

• Artificial neural networks are effective tools for Parkinson's disease diagnosis using voice data.
• Feature selection, particularly using Kendall's correlation coefficient, can enhance diagnostic accuracy.
• Voice analysis combined with ANNs presents a promising avenue for non-invasive PD detection.