Semi-Supervised Learning Based on Competitive Prototype Seeds Selection for Phonation Detection of Patients with Parkinson's Disease

Detection of phonation impairment is very useful for assessing the pathological progress of patients with Parkinson's disease (PD) at early stages. In this paper, we first categorized the acoustic parameters into several families of jitter, shimmer, harmonic-to-noise, frequency, and nonlinear measures, and analyzed the linear correlation coefficients within each acoustic family. Then, we utilized the principal component analysis to reduce the redundant dimensions for different acoustic parameter families. The dominant projected features were chosen for pattern analysis, based on the statistical significant threshold tested by the Wilcoxon rank-sum test. We proposed a novel semi-supervised learning method to detect phonation voice patterns in PD. The semi-supervised learning algorithm contained the competitive prototype seed selection, K-means optimization, and k-nearest neighbor classifications. The experimental results showed that the semi-supervised learning method was able to provide the recall of 0.825, specificity of 0.85, and overall accuracy of 0.838, which was superior to the results of previous related studies in the literature.

Automated summary

Detection of phonation impairment in patients with Parkinson's disease is important for assessing pathological progress, and a novel semi-supervised learning method was proposed to analyze voice patterns in PD. Experimental results showed that the method achieved high recall, specificity, and overall accuracy, outperforming previous studies in the literature.