Session Information
Date: Tuesday, September 24, 2019
Session Title: Parkinsonisms and Parkinson-Plus
Session Time: 1:45pm-3:15pm
Location: Agora 3 West, Level 3
Objective: To identify machine learning classifiers and to assess the acoustic differences between PD patients and healthy control (HC) and the relations between objective and subjective vocal parameters.
Background: Vocal characteristics associated with Parkinson’s disease (PD) are part of hypokinetic dysarthria and can be acoustically analyzed. Voice analysis is being used to diagnose the presence and progression of different diseases, including PD.
Method: The study included 104 PD patients (30 Females) and 82 HC (47 Females). All participants underwent the Montreal Cognitive Assessment (MoCA), Voice Handicap Index (VHI), Beck Depression Inventory (BDI). PD patients underwent the PD quality of life questionnaire (PDQ8), Hoehn & Yahr scale (H&Y). All participants’ voice quality was assessed using the Grade, Roughness, Breathiness, Asthenia & strain (GRBAS) score. All participants were recorded while performing speech tasks including set of vowel sounds, counting up section, a phonetically balanced text to be read aloud, picture description and a short spontaneous speech section. Acoustic analysis including pitch and Root Mean Score (RMS) was performed. A large set of dynamic and static acoustic features were developed in order to train several machine learning classifiers to measure their performance in detecting PD from speech.
Results: Mean age of PD/HC 67.26±10.26/59.09±9.47 respectively. Mean H&Y 2.5±0.81; Disease duration 9.17±6.29 and PDQ8 10.58±6.32; MoCA, VHI, BDI and GRBAS mean score of PD/HC 22.39±4.90/26.54±2.63; 36.27±30.30/4.48±6.51; 3.79±4.19/0.91±1.40; 0.80±0.94/0.15±0.36 (p<0.001) respectively. Significant difference was noted in RMS MAX between PD patients 14.61±29.04 and HC 5.07±0.96. In the HC group a positive correlation was noted between pitch (1.06±0.13) and GRBAS (0±0; p<0.05) during reading. Using special feature classifiers, ~70% correct classification using single feature on \MPT\ task, and ~80% using combination of features were achieved. Classification accuracy for GRBAS among PD patients was 80%.
Conclusion: The combination of acoustic analysis and subjective vocal assessment can differentiate healthy voice from dysarthric voice of PD. Machine learning may be useful in the early detection of PD.
To cite this abstract in AMA style:
Y. Manor, S. Naor, D. Shpunt, N. Diamant, A. Hillel, A. Ezra, I. Opher, Y. Hauptman, R. Aloni-Lavi, A. Faust-Socher, H. Shabtai, R. Peled, A. Migirov, T. Gurevich. Machine learning classifiers and subjective vocal perception of Parkinson’s disease patients and healthy control [abstract]. Mov Disord. 2019; 34 (suppl 2). https://www.mdsabstracts.org/abstract/machine-learning-classifiers-and-subjective-vocal-perception-of-parkinsons-disease-patients-and-healthy-control/. Accessed September 7, 2024.« Back to 2019 International Congress
MDS Abstracts - https://www.mdsabstracts.org/abstract/machine-learning-classifiers-and-subjective-vocal-perception-of-parkinsons-disease-patients-and-healthy-control/