Effective dysphonia detection using feature dimension reduction and kernel density estimation for patients with...

Area of Science:

• Biomedical Engineering
• Speech Science
• Computational Linguistics

Background:

• Dysphonia detection aids in monitoring Parkinson's disease (PD) progression and assessing severity.
• Vocal measurements of sustained phonations offer insights into phonatory impairment.

Purpose of the Study:

• To apply statistical pattern analysis for detecting dysphonia in Parkinson's disease patients.
• To evaluate the efficacy of different classification methods for voice disorder detection.

Main Methods:

• Feature dimension reduction using Sequential Forward Selection (SFS) and Kernel Principal Component Analysis (KPCA).
• Nonparametric kernel density estimation for approximating class-conditional feature densities.
• Classification using Fisher's Linear Discriminant Analysis (FLDA), Maximum a Posteriori (MAP), and Support Vector Machine (SVM) with radial basis function kernels.

Main Results:

• The MAP classifier achieved 91.8% accuracy in distinguishing voice records.
• MAP classifier demonstrated superior diagnostic performance compared to FLDA and SVM.
• Sensitivity, specificity, and ROC curve area under the curve (AUC) for MAP were 0.986, 0.708, and 0.94, respectively.
• Gender was found to be insensitive to dysphonia detection in this cohort.

Conclusions:

• Statistical pattern analysis, particularly with the MAP classifier, is effective for dysphonia detection in Parkinson's disease.
• The MAP classifier offers a promising tool for objective assessment of voice disorders in PD.
• Sustained phonation analysis can identify subtle vocal changes associated with early-stage Parkinson's disease.