A CNN-based Lower Limb Motion Quality Evaluation Method for Home-based Stroke Rehabilitation

Abstract

Background: With the rapid increase of stroke incidence in recent years worldwide, home-based rehabilitation training has become more needed, especially for remote regions or in developing countries where rehabilitation resources are scarce. Studies have demonstrated that home-based rehabilitation for poststroke patients is essential for reducing the cost as well as for providing efficient rehabilitation. Nevertheless, home-based rehabilitation training requires effective professional support and timely evaluation. Method: In this paper, a home-based rehabilitation quality evaluation method for lower limb training was proposed. The kinematic data of a patient’s lower limb during a set of selected training exercises was captured by a wireless body area sensor network (WBASN). The data was then processed by a convolutional neural network (CNN) based algorithm to classify the rehabilitation training type and to evaluate the training quality. A series of kinematic features were selected for rehabilitation quality scoring. The experiments have been conducted using 26 human participants, including 6 healthy participants and 20 stroke patients at different Brunnstrom recovery stages. Results: An accuracy of 95.3% has been achieved for recognizing the rehabilitation training types and a statistically significant positive correlation has been obtained between the objective scores and the Brunnstrom stages evaluated by the clinicians.