Machine Learning Smart System for Parkinson Disease Classification Using the Voice as a Biomarker

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Digital Health

Background:

• Parkinson disease (PD) diagnosis can be challenging.
• Voice changes are a recognized, yet underutilized, biomarker for PD.
• Objective classification of PD using voice data is needed.

Purpose of the Study:

• To develop and evaluate PD Predict, a machine learning system for Parkinson disease classification.
• To assess the efficacy of using voice as a digital biomarker for PD detection.
• To create a user-friendly system for capturing and analyzing voice data.

Main Methods:

• An original dataset of voice recordings was created from the mPower study.
• Audio features, including mel-frequency cepstral coefficient (MFCC) components, were extracted.
• Two machine learning pipelines were trained and validated using nested cross-validation and a holdout set.

Main Results:

• Machine learning pipelines achieved moderate performance (65-75%) in classifying Parkinson disease.
• The system demonstrated good generalizability to unseen data without overfitting.
• The developed system, PD Predict, is accessible via a web application.

Conclusions:

• PD Predict's architecture is effective for Parkinson disease classification.
• The system's performance is promising, validating voice as a digital biomarker.
• Smartphone microphones are viable tools for capturing disease-specific voice biomarkers.