A Machine Learning Framework for Parkinson's Disease Detection Through Turning Metrics.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neurology

Background:

• Parkinson's disease (PD) diagnosis relies on clinical assessment, often subjective and delayed.
• Objective, quantitative measures are needed for early and accurate PD detection.
• Wearable sensor technology offers potential for continuous patient monitoring.

Purpose of the Study:

• To develop and validate a machine learning framework for automated Parkinson's disease detection.
• To utilize turning metrics from inertial measurement units (IMUs) and pressure sensors for PD diagnosis.
• To establish an unsupervised, objective method for early PD identification.

Main Methods:

• A dataset of 29 individuals (healthy, older adults, PD patients) was used.
• An algorithm automatically detected turning points from sensor data.
• Entropy and smoothness metrics were extracted, and machine learning classifiers (SVM, RF, GB) were trained.
• Chi-Squared feature selection was employed.

Main Results:

• The SVM model with Chi-Squared feature selection achieved a 92.92% F1-score in a 5-second window.
• Random Forest and Gradient Boosting models showed high accuracy (84.58% and 87.08%).
• Angular velocity and acceleration were key features, with Bubble Entropy being highly informative.

Conclusions:

• The automated framework demonstrates high potential for early and accurate Parkinson's disease detection.
• This objective, unsupervised approach can aid clinical diagnosis and patient management.
• The system offers a non-invasive, scalable solution for real-world PD screening.