Analysis of voice as an assisting tool for detection of Parkinson's disease and its subsequent clinical interpretation. (April 2021)

• Record Type: Journal Article
• Authors: Solana-Lavalle, Gabriel; Rosas-Romero, Roberto
• Is Part Of: Biomedical signal processing and control. Volume 66(2021)
• Publication Date: 2021-04
• Subjects: Parkinson's disease detection -- Voice analysis -- Feature subset selection -- Assisting tool for diagnosis; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102415
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: Voice-based analysis for Parkinson's disease detection along with framing information. An improvement, over previous works, is shown, in terms of performance detection and complexity. The methodology is based on feature selection, four classifiers, and principal component analysis. Features, with the highest contribution to Parkinson's disease detection, depend on gender. An analysis of feature variability is important to understand the detection outcome. Abstract: In this work, a voice-based analysis is conducted with the contribution of providing physicians with a decision tool along with framing information to help them see functional differences and understand why the detection method suspects PD. The voice-based detection method consists in applying feature subset selection and four different classifiers to voice recordings from five datasets (gender-based, balanced and unbalanced) derived from the largest public dataset for voice-based PD detection. One of the contributions is an improvement over previous works on voice-based PD detection over the same dataset, in terms of performance and complexity. The detection performance is characterized by 95.9% of accuracy, 98.35% of sensitivity, 91.06% of specificity, and 95.6% of precision in women; and 94.36% of accuracy, 100% of sensitivity, 97.1% of specificity, and 96.83% of precision in men. The number of features, fed to classifiers, ranges from 6 to 20. This work shows that different factors are associated with PD detection according to gender: high-frequency voice content is the most significant functional information to assist PD detection in women, while low-frequency content assists PD detection in men better. It is shown that a comparison of the variability of the most important features between patients with PD and controls can be used as contextual information by a physician to have a better interpretation of the classification.