Visualising and quantifying relevant parkinsonian gait patterns using 3D convolutional network

Area of Science:

• Biomedical Engineering
• Computer Vision
• Neurology

Background:

• Parkinson's disease (PD) diagnosis relies heavily on observing motor patterns, particularly gait.
• Current gait analysis often uses marker-based protocols, which can interfere with natural movement.
• Objective, markerless methods are needed for accurate PD progression assessment.

Purpose of the Study:

• To develop an automated, markerless 3D convolutional gait representation for Parkinson's disease classification.
• To create an explainable AI model capable of identifying spatio-temporal gait abnormalities from video data.
• To visualize and interpret the gait patterns contributing to PD classification.

Main Methods:

• A 3D spatio-temporal convolutional network was trained on raw videos and optical flow fields.
• The model was designed to classify markerless gait sequences for PD detection.
• Saliency maps were generated using back-propagation techniques to highlight abnormal motion patterns.

Main Results:

• The approach achieved an average accuracy of 94.89% in classifying Parkinson's disease from gait.
• Saliency maps identified lower limb abnormalities (step length, speed) in PD patients.
• Control subjects showed saliency maps highlighting head and trunk posture variations.

Conclusions:

• Markerless 3D convolutional gait analysis offers a promising, explainable method for PD detection.
• The model effectively identifies key gait features associated with Parkinson's disease.
• This technology can aid in objective disease characterization and treatment monitoring.