A Convolutional Neural Network Based Classification Method for Mild to Moderate Parkinson’s Disease at Turns

Abstract

As the second most common neurodegenerative disease, more than 10 million people live with Parkinson’s disease (PD) and suffer with non-motor and motor symptoms. It is essential to identify PD at early stage for timely treatment. Although early PD may exhibit various but subtle symptoms, they would show more deteriorated gait performance in complex gait tasks, such as turning. In this study, we proposed a convolutional neural network (CNN) based classification model at turns to identify mild and moderate PD from healthy controls. The data derived from sensors attached to the lower limb segments was segmented and normalized into two datasets based on gait cycles and gait tasks (straight walking and turning). Multidimensional data was visualized into RGB images as inputs of CNN which was built with the use of the leave-one-out methods. Results demonstrated that the CNN model at turns obtained a highest accuracy of 91.07% compared to that from straight walking indicating that PD patients exhibited more distinctive motor features during complex gait task. The CNN-based PD multiclassification model is more sensitive to mild and moderate PD patients than the machine learning based models. The proposed method provides a potential tool to identify mild to moderate PD patients in clinical practice.

Xinge Li and Xiayu Huang authors have contributed equally to this work and share first authorship.