Objective: This study aimed to address this gap by investigating whether convolutional neural networks (CNN) can accurately distinguish PD patients from non-PD individuals.

Background: Vocal disorders are strongly associated with early-stage symptoms in 90% of Parkinson’s disease (PD) patients. Detecting voice changes in PD patients could enable early detection and intervention before debilitating physical symptoms appear. Still, research on the ability to distinguish PD patients from other non-Parkinsonian patients (non-PD) is insufficient.

Method: We conducted recordings for a total of 8 tasks, including vowel pronunciation, picture description, procedural narration, and text reading. A total of 163 participants, including 49 PD patients and 114 non-PD individuals (mild cognitive impairment or Alzheimer’s dementia) participated in the voice collection with written consent. The voice recordings were analyzed using Wav2vec 2.0 model which is trained on unlabeled data and learns to extract features or representations from that data without any external guidance.

Results: To correct variables such as age and gender in recruited patients, 49 PD and 49 non-PD patients were analyzed using propensity score matching. Our Wav2vec 2.0 model achieved an accuracy of 92.86%, recall of 93.88%, and F1-score of 92.93%.

Conclusion: This study demonstrates that PD can be distinguished from other non-PD diseases by voice analysis using convolutional neural networks (CNN). Voice analysis can serve as a digital biomarker to facilitate early diagnosis of PD.

« Back to 2023 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/automatic-detection-of-parkinsons-disease-by-voice-analysis/