Dysarthria Detection with Deep Representation Learning for Patients with Parkinson's Disease

Dysarthria is a very common motor speech symptom in Parkinson's disease impairing normal communications of patients. Detection of dysarthria could assist clinical diagnosis and intervention of Parkinson's disease, provide monitoring approach for treatment-related side effects, and lead to effective speech therapy to prevent further communication and social deficits. Applying machine learning techniques to speech analysis for patients offers more resource-efficient, accessible and objective tools for screening and assessment of dysarthria. In this study, we constructed a multi-task speech dataset recorded from 600 participants with high data diversity to facilitate the development of detection models. We established a remote data acquisition and end-to-end prediction pipeline with deep representation learning, compared the performance with different feature-based learning and classification methods, and achieved a superior accuracy of over 90%. Different affecting factors were analyzed for model performance. Our proposed framework demonstrates the potential of developing and deploying an automated self-monitoring approach of dysarthria for patients with Parkinson's disease, which could benefit a large-scale population and their disease managements.