Classification of Gait Abnormality Associated with Parkinson'S Disease Using Deep Learning Neural Networks with Inputs of Foot Pressures

Detection and classification of gait abnormality is crucial for early detection, gait training and cuing, and medicine guiding in persons with Parkinson’s disease (PD). Previous gait abnormality or gait parameters for PD subjects such as freeze of gait are mainly detected using lower-extremity accelerations which are less related to dynamic information of foot pressures. Recent applications of foot pressures are most used for the investigation force distributions of feet in subjects of diabetes and cerebral palsy but less for detecting gait abnormality in PD subjects. In particular, the aforementioned approaches are based on the hand-crafted features from foot pressures. However, the gait abnormality in PD subjects, such as bradykinesia, festinating gait, shuffling steps and freeze of gait, involves spatial-temporal changes of foot pressures, whereby needs techniques of dynamic pattern recognition. Hence, we propose a concept of dynamic images processing on foot pressures in this proposal. Convolutional neural networks and recurrent neural networks will be employed to classify bradykinesia, festinating gait, shuffling steps, freeze of gait and normal gait based on the collected foot pressure data in clinics. The aforementioned works will be executed in two years including a Cortex-based device with full-placed foot pressure sensing, data visualization and machine learning architectures in the first year; and an APP in smartphone with foot pressure visualization and deep learning for detecting gait abnormality in PD subjects in the second year.

Project ID:PB10901-2114
External Project ID:MOST108-2221-E182-013-MY2