Classification of Parkinson's disease from smartphone recording data using time-frequency analysis and convolutional...

Area of Science:

• Neurology
• Biomedical Engineering
• Data Science

Background:

• Parkinson's disease (PD) is a neurodegenerative disorder diagnosed via clinical observation.
• Voice impairment is a common symptom affecting most PD patients.
• Current diagnostic methods rely on specialist expertise.

Purpose of the Study:

• To evaluate smartphone voice recordings as a non-invasive tool for PD diagnosis and monitoring.
• To assess the feasibility of using one-second voice samples for self-diagnosis.
• To analyze data from the mPower study, a large mobile PD research initiative.

Main Methods:

• Utilized 29,798 ten-second voice recordings from 4,051 participants in the mPower study.
• Generated 385,143 one-second audio samples from sustained phonation (/aa/).
• Applied Convolutional Neural Network (CNN) models to spectrograms of audio samples for classification.

Main Results:

• Achieved high classification accuracies using CNN models: LeNet-5 (97.7%), ResNet-50 (98.6%), and VGGNet-16 (99.3%).
• Demonstrated robust performance with a generalized approach on a large dataset.
• Highlighted the effectiveness of one-second audio clips.

Conclusions:

• One-second smartphone voice recordings show promise as a non-invasive biomarker for Parkinson's disease.
• This method offers a potential avenue for remote PD diagnosis and disease progression monitoring.
• The findings support the development of smartphone-based self-diagnosis tools for PD.