An EMG signal driven Wheelchair controller for Spinal Cord Injury(SCI) Patients

Abstract

In this article, the use of an EMG control based wheelchair is presented. The me- chanical design is based on six gestures and eight channel EMG sensors that are controlled by six hand directions. Electric EMG-based wheelchairs are essential for SCI patients. Because cervical spinal cord injury (SCI) creates severe sensory anomalies in their bodies, SCI patients are unable to walk. Their hand nerves do not respond correctly due to their upper limb impairments. It takes a long time for their hand to react. As a result, SCI patients are unable to drive a normal joystick wheelchair manually. They must rely on others to utilize these sorts of wheelchairs. This study attempts to give a novel alternative controlling technol- ogy for people with paralysis, particularly of the feet and hands, by creating an electric wheelchair that uses the control of electromyography (EMG) signal as a consequence of muscle relaxation. EMG signals are commonly utilized to quantify torques in human muscles. We have created an electromyography (EMG)-based hand gesture dataset to control electric wheelchair for the patient with spinal cord injury (SCI). We have recorded eight-channel surface EMG (sEMG) signals from the EMG sensor placed at the forearm of the SCI patient. These signals were collected from six hand gesture-based wheelchair control movements (for- ward, reverse, turn left, turn right, start and stop). We collected hand gesture data containing different EMG signals from 12 healthy subjects and 7 SCI sub- jects. Later on, The EMG signals were segmented and the time-domain feature extraction technique was applied to generate 18000 training samples and 10500 testing samples. We then classified the hand gestural EMG signals using 5 differ- ent classical machine learning models. We analyze the classification results in two ways. The first one is, training the models using only data of healthy subjects and cross-validated using data from 7 SCI patients. And the second one is by in- cluding six SCI patient’s data in the training process along with healthy subjects we performed leave one out cross-validation. From this analysis we were able to achieve the highest 95.42% accuracy using decision tree (DT) and Random Forest (RF) algorithms.