CNN-based Identification of Parkinsonian Gait using Ground Reaction Forces

Parkinson's disease is among the most common conditions faced by older people, which causes stiffness and slowing of movement. One manifestation of the disorder assumes a non-physiological walking pattern characterized by flatfoot strike or toe-to-heel progression. This paper proposes a method to identify Parkinson's disease using Convolutional Neural Networks to classify the ground reaction forces exerted during walking. Bilateral plantar pressure is acquired using three Aidong IMB C20B pressure sensors for each foot. The training database consists of 13 records acquired in clinical environment and an additional 72 records publicly available online in the Gait in Parkinson's Disease database. The signals are pre-processed to obtain 2D images, which are used to train three Convolutional Neural Networks models: MobileNet, ResNet50 and DenseNet121. According to the performance metrics, the results obtained to identify Parkison's disease are remarkable. The best accuracy is 89% and is obtained for the MobileNet model.