Stacked Spatial and Temporal Deep Learning Methods for Identification of Parkinson’s Disease Using Gait Signals
Conference Paper
Overview
Research
Identity
Additional Document Info
View All
Overview
abstract
Parkinson’s disease (PD) is a progressive condition that affects dopaminergic neurons, causing motor alterations. Motor disturbances, such as gait impairment, can be used to assess the disease. Unfortunately, gait disturbances, such as decreased walking speed and step variability, can also occur due to aging, affecting the identification of abnormal PD gait. Therefore, developing an adequate tool to evaluate PD patients’ gait is essential. This paper proposes a deep learning algorithm to differentiate between PD gaits and normal walking using vertical ground reaction force (VGRF) signals. CLDNN is a single framework composed of a convolutional neural network, a long-short term memory network, and a deep neural network. To train and validate a CLDNN classifier gait cycles were obtained from VGRF signals. The VGRF signals were from a public database with recordings from 93 PD patients and 73 healthy adult controls. The CLDNN performance was evaluated by five-fold cross-validation. The combined spatial and temporal methods in CLDNN enabled the effective identification of PD gait with less complex architecture. The best weighted accuracy was 98.28 ± 0.38. Thus, our model is compact and efficient for future embedded or portable implementations.
