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

Abstract

Parkinson's disease (PD) is a common neurodegenerative disorder worldwide, with over 6.2 million registered cases. Gait analysis plays a fundamental role in evaluating motor abnormalities associated with this disease. However, current methods, such as marker-based systems, are intrusive and expert-dependent. Markerless alternatives, like video sequence analysis, have been proposed, but they tend to provide overall classification scores and lack the ability to interpret joint kinematics in detail. An innovative technique is presented using volumetric convolutional networks that can learn intermediate postural patterns and distinguish between Parkinson's patients and control subjects. This approach utilizes OpenPose activations and then applies hierarchical convolution to minimize classification. In tests conducted with 14 Parkinson's patients and 16 control subjects, this method achieved a classification accuracy of 98%.