Gait classification for Parkinson's Disease using Stacked 2D and 1D Convolutional Neural Network

Abstract

Parkinson's Disease (PD) affects a significant amount of elderlies around the world. The progression and effectiveness of medication can be inferred from changes or improvements of gait in patients. Vertical ground reaction force is a common used measure to classify gait. In this study, such data from a publicly available dataset is used to classify gait between PD patients and healthy control subjects of similar ages. The data were preprocessed by normalization, splitting into standard time units and then converted to images. Classification was done using a stacked 2-dimensional and 1-dimenisonal convolutional neural network. Our model achieved 88.7% accuracy, an improvement of 5.3% over the next best performing model that was published.