A Volumetric Deep Architecture to Discriminate Parkinsonian Patterns from Intermediate Pose Representations

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Parkinson's disease (PD) affects over 6.2 million people globally, with gait abnormalities being a key diagnostic indicator.
• Current gait analysis methods (e.g., marker-based) are often intrusive and require expert interpretation.
• Existing markerless video analysis techniques provide limited kinematic detail and overall classification scores.

Purpose of the Study:

• To develop an innovative, non-intrusive method for analyzing gait in Parkinson's disease patients.
• To utilize volumetric convolutional networks for detailed joint kinematic interpretation.
• To accurately differentiate Parkinson's disease patients from healthy controls using AI.

Main Methods:

• Implementation of volumetric convolutional networks to learn intermediate postural patterns from video data.
• Utilizing OpenPose activations as input for the network.
• Application of hierarchical convolution for minimizing classification errors.

Main Results:

• The developed method achieved a high classification accuracy of 98% in distinguishing Parkinson's patients from control subjects.
• The technique successfully learned intermediate postural patterns indicative of PD.
• Demonstrated the ability to interpret joint kinematics in detail, surpassing previous markerless methods.

Conclusions:

• Volumetric convolutional networks offer a promising, non-intrusive approach for Parkinson's disease diagnosis through gait analysis.
• This AI-driven method provides detailed kinematic insights, overcoming limitations of current techniques.
• The high accuracy suggests significant potential for clinical application in early PD detection and monitoring.