Classification of Dysphonic Voices in Parkinson's Disease with Semi-Supervised Competitive Learning Algorithm.

Area of Science:

• Biomedical Engineering
• Computational Linguistics
• Neurology

Background:

• Parkinson's disease (PD) diagnosis can be challenging, necessitating advanced analytical tools.
• Vocal pattern analysis offers a non-invasive method for PD detection.
• Existing classification methods may not fully capture the complexity of vocal biomarkers.

Purpose of the Study:

• To introduce a novel semi-supervised competitive learning (SSCL) algorithm for classifying vocal patterns in Parkinson's disease (PD).
• To reduce acoustic parameter dimensionality using Principal Component Analysis (PCA) and identify significant features for PD detection.
• To evaluate the diagnostic performance of the proposed SSCL algorithm against conventional classifiers.

Main Methods:

• Acoustic parameters (jitter, shimmer, harmonic-to-noise, frequency, nonlinear measures) were analyzed for linear correlations.
• Principal Component Analysis (PCA) was employed to reduce data dimensionality, followed by Mann−Whitney−Wilcoxon tests for feature significance (p < 0.05).
• The SSCL algorithm integrated competitive prototype seed selection, K-means optimization, and nearest neighbor classification.

Main Results:

• High correlations were found among jitter, shimmer, and harmonic-to-noise parameters.
• Eight dominant PCA-projected features were selected based on eigenvalue thresholds and statistical significance.
• The SSCL algorithm achieved high diagnostic performance metrics: accuracy (0.838), recall (0.825), specificity (0.85), precision (0.846), F-score (0.835), MCC (0.675), AUC (0.939), and Kappa (0.675).

Conclusions:

• The proposed SSCL algorithm demonstrates superior performance in classifying vocal patterns for Parkinson's disease detection compared to KNN and SVM.
• The study highlights the potential of SSCL combined with PCA for developing accurate and efficient PD diagnostic tools.
• Vocal analysis using advanced machine learning techniques offers a promising avenue for non-invasive PD screening and diagnosis.