Robust parsimonious selection of dysphonia measures for telemonitoring of Parkinson's disease symptom severity

(UPDRS) which spans the range 0-176 (0 denotes healthy). This assessment requires the patient's physical presence in the clinic, is time consuming, and relies on the clinical rater's subjective evaluation and experience; practice has shown that expert clinicians might differ by as much as 4-5 UPDRS points in their evaluations. We had previously developed a statistical machine learning framework which enables accurate and objective quantification of average PD symptom severity using exclusively speech signals. For this purpose, we evaluated 132 speech signal processing algorithms (dysphonia measures), which attempt to capture distinctive characteristics in PD subjects' voice. On a very large database of about 6,000 phonations, we could replicate the clinical experts' assessments within less than two UPDRS points' error. In this paper, we focus on identifying the most successful of the original 132 dysphonia measures in estimating UPDRS using five robust feature selection techniques. We demonstrate that we can improve on our previous findings using only 15 dysphonia measures, where the selected measures also tentatively indicate the most representative pathophysiological characteristics in male and female PD voices.