Prediction of Parkinson Disease Using Long-Term, Short-Term Acoustic Features Based on Machine Learning.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Biology

Background:

• Parkinson's disease (PD) is a prevalent neurodegenerative disorder with both motor and non-motor symptoms.
• A prodromal phase, characterized by non-motor symptoms like sleep disturbances and voice changes, often precedes overt motor symptoms.
• Speech analysis shows promise as a digital biomarker for early PD detection and patient monitoring.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) models in detecting Parkinson's disease (PD) through voice analysis.
• To identify the most effective acoustic features and ML algorithms for distinguishing PD patients from healthy subjects.
• To establish voice analysis as a non-invasive tool for early PD diagnosis and prognosis.

Main Methods:

• A cross-sectional study analyzed voice impairments in 40 PD patients and 41 healthy individuals.
• Utilized a diverse set of acoustic features: long-term (jitter, shimmer, CPP), short-term (MFCC), and non-standard (PPE, RPDE).
• Trained and evaluated multiple ML algorithms (RF, KNN, DT, NB, SVM, LR) using cross-validation for performance assessment.

Main Results:

• The random forest (RF) model achieved the highest performance, with 82.72% accuracy and an 89.65% ROC-AUC score.
• Support vector machine (SVM) also showed considerable efficacy (75.29% accuracy, 82.63% ROC-AUC).
• Combining a comprehensive set of acoustic features enhanced predictive performance compared to studies using limited feature subsets.

Conclusions:

• Advanced acoustic analysis coupled with ML algorithms offers a reliable, non-invasive method for early PD detection.
• This approach holds significant potential for improving healthcare outcomes through timely diagnosis and patient management.
• Voice analysis can serve as a valuable digital biomarker for early identification and follow-up of Parkinson's disease.