Goal:Human voice is a promising noninvasive indicator for diagnosing Parkinson’s Disease (PD). It is also unique since it can be collected remotely, increasing accessibility to a wide range of underprivileged patients. However, recognizing PD’s signature in the human voice is nontrivial since the available features are many, and the signal may be noisy. Methods: A new mechanism based on Interval Type-2 Fuzzy Analytical Hierarchy Process is proposed here for choosing a reduced feature set from 339 dysphonia speech features, based on five criteria of 1) Robustness, 2) Relief, 3) Minimum Redundancy and Maximum Relevance, 4) Gaussian mixture model separation, and 5) Classifier separation ability. A Least Squares Support Vector Machine then categorizes the samples as belonging to either a healthy subject or a patient with PD. The database of 47 subjects with an average age of 67 is obtained from the elderly in nursing homes and Parkinson’s specialized clinics. By reducing signal quality similar to a standard phone line, we study the teleoperation prospect of the proposed technique. Results: Ten-fold cross-validation shows an overall accuracy of 95.32%(93.11%) for noiseless(noisy) conditions, with separate analysis for male, female, and both genders populations. Furthermore, Leave-One-Speaker-Out analysis yields an overall accuracy of 93.11%(84.61%) for noiseless(noisy) conditions. Conclusion: The proposed strategy offers viable remote PD diagnosis with higher accuracy for the male population. Significance: The proposed method suggests reduced feature sets that meet differing objectives of simplicity, performance, and robustness. Results could be particularly significant in PD diagnosis in remote areas.

Keywords