Parkinson disease detection based on in-air dynamics feature extraction and selection using machine learning.

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Parkinson's disease (PD) diagnosis often relies on broad statistical features from handwriting, missing crucial dynamic movement details.
• Traditional methods struggle with accuracy, robustness, and sensitivity due to oversimplified feature representation and lack of focus on subtle variations.

Purpose of the Study:

• To develop an optimized Parkinson's disease detection methodology using novel dynamic kinematic features and machine learning.
• To classify Parkinson's disease patients into distinct stages (early, mid, late) based on disease progression, moving beyond simple PD vs. non-PD differentiation.

Main Methods:

• Extracted 65 novel and 23 existing kinematic features focusing on acceleration, deceleration, and directional changes during handwriting.
• Applied statistical formulas for hierarchical feature enhancement and Sequential Forward Floating Selection for feature optimization.
• Utilized an ensemble machine learning approach with voting for classification.

Main Results:

• Achieved 96.99% accuracy in task-wise classification and 99.98% accuracy in task ensembles on the PaHaW dataset.
• Surpassed the existing state-of-the-art model by 2% in accuracy.
• Demonstrated superior performance in capturing subtle movement variations indicative of Parkinson's disease.

Conclusions:

• The proposed methodology significantly enhances Parkinson's disease detection accuracy and staging capabilities.
• Incorporating dynamic kinematic features and advanced machine learning offers a more sensitive and robust approach to PD assessment.
• This work sets a new benchmark for Parkinson's disease detection using handwriting analysis.