Dynamic Modeling Of Parkinsonian Gait Using Latent Biomarkers

Abstract

Parkinson's Disease (PD) Affects Gait Dynamics, And Hence, Objective Analysis Is Required For Accurate Diagnosis. In This Regard, This Manuscript Proposes A Deep Learning Framework For Modeling Parkinsonian Gait Using Vertical Ground Reaction Force (VGRF) Signals. A Hybrid CNN-LSTM Network Is Used To Effectively Capture Spatial And Temporal Features Of The Gait, And A 16-dimensional Latent Space Is Used To Effectively Capture Discriminative Gait Features. An Accuracy Of 99%, Along With High Precision And Recall, Is Achieved By The Network, And A High AUC Of 0.999 Indicates Effective Separability Of Classes. Furthermore, Using Principal Component Analysis On The Learned 16D Space, Distinct Clusters Of Healthy And Parkinsonian Gait Patterns Are Observed. It Is Thus Concluded That The Proposed Framework Is Effective For Accurate Detection Of PD Using Gait Analysis.

Keywords