Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Early Recognition of Parkinson's Disease Through Acoustic Analysis and Machine Learning

Niloofar Fadavi, Nazanin Fadavi

TL;DR

This work surveys acoustic-based Parkinson's disease recognition using machine learning, emphasizing data wrangling of a 23-feature voice dataset and the evaluation of multiple classifiers. Across traditional and ensemble methods, Generalized Forest, SVM, and gradient boosting approaches deliver the strongest performance (around 0.9487 accuracy and 0.9412 precision), while PCA and feature selection influence some models differently. Neural networks offer potential gains but require more data and computing resources, with feedforward nets outperforming certain recurrent architectures in this study. The findings underscore the potential of noninvasive speech analysis for early PD detection and outline directions like leveraging raw sequential voice data with RNNs for future improvements.

Abstract

Parkinson's Disease (PD) is a progressive neurodegenerative disorder that significantly impacts both motor and non-motor functions, including speech. Early and accurate recognition of PD through speech analysis can greatly enhance patient outcomes by enabling timely intervention. This paper provides a comprehensive review of methods for PD recognition using speech data, highlighting advances in machine learning and data-driven approaches. We discuss the process of data wrangling, including data collection, cleaning, transformation, and exploratory data analysis, to prepare the dataset for machine learning applications. Various classification algorithms are explored, including logistic regression, SVM, and neural networks, with and without feature selection. Each method is evaluated based on accuracy, precision, and training time. Our findings indicate that specific acoustic features and advanced machine-learning techniques can effectively differentiate between individuals with PD and healthy controls. The study concludes with a comparison of the different models, identifying the most effective approaches for PD recognition, and suggesting potential directions for future research.

Early Recognition of Parkinson's Disease Through Acoustic Analysis and Machine Learning

TL;DR

This work surveys acoustic-based Parkinson's disease recognition using machine learning, emphasizing data wrangling of a 23-feature voice dataset and the evaluation of multiple classifiers. Across traditional and ensemble methods, Generalized Forest, SVM, and gradient boosting approaches deliver the strongest performance (around 0.9487 accuracy and 0.9412 precision), while PCA and feature selection influence some models differently. Neural networks offer potential gains but require more data and computing resources, with feedforward nets outperforming certain recurrent architectures in this study. The findings underscore the potential of noninvasive speech analysis for early PD detection and outline directions like leveraging raw sequential voice data with RNNs for future improvements.

Abstract

Parkinson's Disease (PD) is a progressive neurodegenerative disorder that significantly impacts both motor and non-motor functions, including speech. Early and accurate recognition of PD through speech analysis can greatly enhance patient outcomes by enabling timely intervention. This paper provides a comprehensive review of methods for PD recognition using speech data, highlighting advances in machine learning and data-driven approaches. We discuss the process of data wrangling, including data collection, cleaning, transformation, and exploratory data analysis, to prepare the dataset for machine learning applications. Various classification algorithms are explored, including logistic regression, SVM, and neural networks, with and without feature selection. Each method is evaluated based on accuracy, precision, and training time. Our findings indicate that specific acoustic features and advanced machine-learning techniques can effectively differentiate between individuals with PD and healthy controls. The study concludes with a comparison of the different models, identifying the most effective approaches for PD recognition, and suggesting potential directions for future research.
Paper Structure (17 sections, 1 figure, 11 tables)

This paper contains 17 sections, 1 figure, 11 tables.

Table of Contents

  1. Introduction
  2. Data Wrangling
  3. Summary Statistics
  4. Analysis of Pairwise Correlation Heatmap
  5. Alternative Methods for PD Classification
  6. Generalized Forest
  7. Logistic Regression
  8. Support Vector Machines
  9. Gradient Boosting Machines
  10. XGBoost
  11. LightGBM
  12. CatBoost
  13. Results
  14. Neural Networks
  15. Feedforward Neural Networks
  16. ...and 2 more sections

Figures (1)

  • Figure 1: Heat-Map