Dysphonic Voice Pattern Analysis of Patients in Parkinson's Disease Using Minimum Interclass Probability Risk Feature...

Area of Science:

• Speech analysis
• Bioacoustics
• Medical diagnostics

Background:

• Dysphonia and phonation disorders significantly impact communication.
• Accurate detection and assessment of voice disorders rely on quantitative voice analysis.
• Existing feature selection methods may not optimally identify critical vocal parameters.

Purpose of the Study:

• To develop and evaluate a novel feature selection method for dysphonic voice detection.
• To compare the proposed interclass probability risk (ICPR) method with modified Kullback-Leibler divergence (MKLD).
• To assess the performance of various classification algorithms using selected vocal features.

Main Methods:

• Analysis of linear correlations between 22 voice parameters.
• Application of linear discriminant analysis to combine correlated parameters.
• Estimation of probability density functions using the Parzen-window technique.
• Development of the interclass probability risk (ICPR) method for feature selection.
• Comparison with the modified Kullback-Leibler divergence (MKLD) approach.
• Classification using generalized logistic regression analysis (GLRA), support vector machine (SVM), and Bagging ensemble algorithms.

Main Results:

• The ICPR method identified dominant vocal features more effectively than MKLD.
• Classifiers utilizing ICPR features demonstrated superior classification performance.
• Support vector machine (SVM) achieved a specificity of 0.8542 in distinguishing normal vocal patterns.
• The Bagging ensemble algorithm with ICPR features achieved the highest sensitivity (0.9796) and area under the ROC curve (0.9558).

Conclusions:

• The proposed ICPR feature selection method is effective for dysphonic voice detection.
• Machine learning algorithms, particularly the Bagging ensemble, perform well with ICPR-selected features.
• This approach enhances the accuracy and reliability of voice disorder assessment.