Detecting Parkinson Disease Using a Web-Based Speech Task: Observational Study

Area of Science:

• Neurology
• Computational Linguistics
• Biomedical Engineering

Background:

• Parkinson disease (PD) diagnosis and care access are limited globally, especially in resource-poor nations.
• Rising PD cases are projected worldwide due to aging populations and environmental factors.
• Timely PD diagnosis is crucial for effective medical intervention and improved patient quality of life.

Purpose of the Study:

• To introduce a web-based framework for remote Parkinson disease (PD) screening using speech analysis.
• To enable individuals worldwide to record speech data for PD screening.

Main Methods:

• Collected speech data from 726 participants (PD and non-PD).
• Extracted acoustic and deep learning features from speech samples, including mel-frequency cepstral coefficients and dysphonia features.
• Trained machine learning models, such as XGBoost, and used Shapley additive explanations for feature importance.

Main Results:

• Achieved an area under the curve of 0.753 for PD detection using acoustic features and XGBoost.
• Mel-frequency cepstral coefficients and specific dysphonia features were most influential in model decisions.
• The model demonstrated consistent performance across laboratory and real-world data, irrespective of gender or age.

Conclusions:

• The developed tool facilitates remote PD screening via audio-enabled devices, enhancing access to neurological care.
• This approach contributes to greater equity in healthcare by overcoming geographical and resource limitations.
• The framework supports widespread data collection for improved PD detection and management.