Voice Analysis with Machine Learning: One Step Closer to an Objective Diagnosis of Essential Tremor

Abstract

A BS TRACT: Background: Patients with essential tremor have upper limb postural and action tremor often associated with voice tremor. The objective of this study was to objectively examine voice tremor and its response to symptomatic pharmacological treatment in patients with essential tremor using voice analysis consisting of power spectral analysis and machine learning. Methods: We investigated 58 patients (24 men; mean age ± SD, 71.7 ± 9.2 years; range, 38-85 years) and 74 age-and sex-matched healthy subjects (20 men; mean age ± SD, 71.0 ± 12.4 years; range, 43-95 years). We recorded voice samples during sustained vowel emission using a high-definition audio recorder. Voice samples underwent sound signal analysis, including power spectral analysis and support vector machine classification. We compared voice recordings in patients with essential tremor who did and did not manifest clinically overt voice tremor and in patients who were and were not under the symptomatic effect of the best medical treatment.Results: Power spectral analysis demonstrated a prominent oscillatory activity peak at 2-6 Hz in patients who manifested a clinically overt voice tremor. Voice analysis with support vector machine classifier objectively discriminated with high accuracy between controls and patients who did and did not manifest clinically overt voice tremor and between patients who were and were not under the symptomatic effect of the best medical treatment. Conclusions: In patients with essential tremor, voice tremor is characterized by abnormal oscillatory activity at 2-6 Hz. Voice analysis, including power spectral analysis and support vector machine classification, objectively detected voice tremor and its response to symptomatic pharmacological treatment in patients with essential tremor.

Reproducibility of Voice Analysis with Machine Learning

“…Indeed, a recent study demonstrated that the repeatability of automatically extracted speech features for dysarthria assessment using OpenSMILE is low. 4 Therefore, to demonstrate the impropriety of such a design, we performed two additional experiments to challenge the original analysis by Suppa et al 1 First, we generated a random vector of values ranging between 0 and 1 to substitute 6139 hypothetical voice features across 58 hypothetical patients. Second, we used OpenSMILE 3 to extract 6139 voice features via sustained phonation /a/ paradigm recorded using standardized settings 5 from two groups of healthy speakers designed to be age-and gender matched to the original ET VT+ and ET VT-groups (Fig.…”

“…1). 1 For both scenarios, we performed a similar procedure as reported in Suppa et al 1 ; in short, the generated/extracted features underwent feature selection, and 20 best-ranked features were used to train a linear kernel support vector machine classifier with a 10-fold cross-validation procedure.…”

Fostering Voice Objective Analysis in Patients with Movement Disorders

“…[3][4][5] We therefore suggest that the sustained emission of a vowel could be the most appropriate technical solutions for preventing linguistic confounding and achieving standardized worldwide procedures. [3][4][5] Finally, the recording guidelines of Rusz et al 1 require voice samples collected only in a specialized laboratory using dedicated hi-tech audio recorders under expert supervision. We have recently proposed an alternative solution consisting of voice recordings acquired by currently available smartphones that can simplify the experimental procedures in a more ecologic scenario and provide the background for future application in the context of telemedicine.…”

“…1 However, recent data from our group show similar results when comparing the sustained emission of vowels and sentences (eg, connected speech). [3][4][5] We therefore suggest that the sustained emission of a vowel could be the most appropriate technical solutions for preventing linguistic confounding and achieving standardized worldwide procedures. [3][4][5] Finally, the recording guidelines of Rusz et al 1 require voice samples collected only in a specialized laboratory using dedicated hi-tech audio recorders under expert supervision.…”

Reply to: “Reproducibility of Voice Analysis with Machine Learning”

“…We hope that the scientific impact of our first analysis is now much more clear. 2 The second point raised by Rusz et al 1 concerns our tool kit for voice feature extraction. We believe that the tool kit Receiver operating characteristic curves calculated with a support vector machine using 20 best-ranked features selected from 6139 features to differentiate two groups (34 and 24 samples, respectively) formed by (A) random values ranging from 0 to 1 and (B) healthy speakers with voice analysis based on sustained phonation paradigm extracted via OpenSMILE toolkit.…”