Machine learning for early detection and severity classification in people with Parkinson's disease

Area of Science:

• Neuroscience
• Biomedical Engineering
• Clinical Neurology

Background:

• Early detection and accurate staging of Parkinson's disease (PD) are crucial for effective treatment and rehabilitation.
• Current methods for early PD detection and motor symptom severity classification using gait analysis lack consensus.
• Gait analysis offers a non-invasive approach to assess motor function in neurodegenerative diseases.

Purpose of the Study:

• To evaluate the accuracy of machine learning models in classifying early and moderate-stage Parkinson's disease (PD) using spatiotemporal gait features.
• To determine the effectiveness of different walking speeds (preferred, faster, slower) in gait-based PD detection and staging.
• To identify key gait parameters that best differentiate PD stages and healthy controls.

Main Methods:

• Recruited 178 participants: 103 with PD (61 early-stage, 42 moderate-stage) and 75 healthy controls.
• Collected spatiotemporal gait data on a 24-m walkway at preferred walking speed (PWS), 20% faster (HWS), and 20% slower (LWS).
• Applied machine learning algorithms (Random Forest, Naïve Bayes) to classify PD stages based on gait features like walking speed, stride length, and coefficient of variation (CV) of stride length.

Main Results:

• Random Forest model achieved 78.1% accuracy in classifying early/moderate PD using PWS walking speed, HWS stride length, and LWS stride length CV.
• Naïve Bayes model achieved 67.3% accuracy for early PD detection using HWS stride length and LWS stride length CV.
• Preferred walking speed (PWS) was the most critical feature for distinguishing early from moderate PD, with 69.8% accuracy.

Conclusions:

• Gait analysis, particularly when assessing variations across different walking speeds, shows promise for early Parkinson's disease detection.
• Specific gait parameters, including walking speed and stride length variability, can aid in classifying PD severity.
• Machine learning models applied to gait data offer a potential tool for objective assessment in Parkinson's disease management.